A local layer of control, savings and safety for coding agents. It measures what every turn costs, cuts the waste and protects your work — no proxy, no MCP, nothing leaving your machine.
A token's cost does not end when the response arrives. It keeps being paid on every following turn that content remains in context:
real cost = token price × turns it survives in context
A 15k-token output that enters on turn 3 of a 40-turn session is sent again 37 times. That is 555k token-turns for a single answer. HarnessUp measures that in your own transcripts — not in a market average — and shows where the money went.
Measured, not promised. The hook answers at a p95 of ~9 ms with a hard 2 s ceiling. Over budget? It returns {} and your agent moves on. HarnessUp is never why a session hangs.
What it is not
Not a plugin, MCP, skill or proxy. It is an installable product with its own daemon, panel, license and account.
It does not intercept your API. It runs the official claude and codex with your account and uses the hooks the harness already offers.
It does not read your token. Never reads, copies, stores or forwards OAuth tokens, API keys or cookies.
It does not guess. Before acting, it calculates. Shadow mode shows the savings that would happen without applying anything.
What changes in your day
Before
With HarnessUp
You find out the spend at the end of the month.
The panel shows usage, cuts and leftovers per session and model.
The agent re-reads the same huge file every turn.
Read Router returns an index on the first read and the full file on the second.
A long session gets expensive without you noticing.
Renew Session points to the turn where restarting costs less than continuing.
A command deletes something it should not.
UltraShield moves it to quarantine; Undo restores the state.
A secret slips into context.
Secret Lock masks the credential before it is sent.
Every path installs the same thing: three programs and nothing else. There is no system service and nothing is written outside your user folder.
Program
What it is
harnessupd
The daemon: measures, runs modules and serves the panel on 127.0.0.1.
harnessup
The command line.
harnessup-hook
The fallback binary the harness calls from hooks.
Windows — one line in PowerShell
The fastest path, and it does not need Node.
irm https://harnessup.dev/install.ps1 | iex
The script downloads the latest zip, checks the SHA-256 against the manifest before unpacking, installs into %LOCALAPPDATA%\HarnessUp\, adds it to PATH and connects Claude Code. If the hash does not match, nothing is written.
Open a new terminal after installing so harnessup exists on PATH.
Windows — installer
Download harnessup-<version>-windows-x64-setup.exe from the download page and run it. It installs per user, does not ask for admin and puts everything in %LOCALAPPDATA%\HarnessUp\. To remove it, use “Apps and Features”: hooks are removed before the files.
Windows and macOS — npm
npm install -g harnessup
Requires Node 18 or newer. The package downloads only the binary for your platform; there is no install script fetching things from the network, so it works the same on machines using --ignore-scripts. npm installs do not self-update: run npm install -g harnessup@latest.
macOS and Linux — curl
curl -fsSL https://harnessup.dev/install.sh | sh
The script checks the hash, installs into your user folder and puts the command on PATH. On macOS this is the recommended path. On Linux, support is outside v1 and install.sh has been validated on Ubuntu 24.04.
Check the file before running it
Every published zip has a SHA-256. Compare it with SHA256SUMS.txt:
The Windows warning is expected and correct. The executable has no code-signing certificate yet, so SmartScreen does not know who published it. Check the hash first and only then proceed. A mismatched hash is the only case where you should delete the file. See the warning page.
Connect Claude Code and Codex
Run once to write the hooks into Claude Code's configuration, install the /up command and arm the channel test:
harnessup install
Then start the daemon:
harnessupd
If you installed from the zip, call both with a path: .\harnessup.exe install and .\harnessupd.exe.
Codex: hooks are installed, but Codex asks for confirmation. Open /hooks in Codex and mark the HarnessUp hooks as trusted.
Check that it works
harnessup doctor
The doctor answers what nothing else does: whether hooks are installed, which events actually arrived — registered is not the same as called — and which channels the product can use to reach you on this machine.
If [speech channels] says not tested, open Claude Code, send any message and type the code that appears:
harnessup channels <code>
Update and uninstall
Installer, irm and zip: do nothing. The daemon downloads in the background, checks signature and hash, and swaps the files on its next shutdown — never mid-session. The panel announces new versions on every channel.
To remove:
harnessup uninstall
This removes hooks, /up and leaves the machine as it was. Measured data stays in ~/.harnessup/; delete the folder if you want it all gone.
Start here · 5 min read
Quickstart
From install to the first diagnosis in five steps. Measurement is free and needs no login; a license only enters when you want the paid modules.
Install and check the integration
Run the installer for your system, then harnessup install and harnessup doctor. The doctor must show hooks installed and events arriving.
Open the panel
harnessup open opens the UI at 127.0.0.1:7717. It is local, with a per-install access token.
Run the first measurement
Deep Scan reads the local history and shows the waste in sessions you already have. Nothing is sent out.
Turn on shadow mode
See how much HarnessUp would have cut without applying anything. It is the honest way to decide.
Plug in the license
When it makes sense, activate the key in the panel, the CLI or the link in your purchase email. The license floats: unplug one machine and use another.
The first commands
# what has already been measured
harnessup scan
harnessup report
# simulate without applying
harnessup shadow on
harnessup shadow status
# list and enable modules
harnessup modules list
harnessup modules enable secret-lock
# license
harnessup license status
harnessup license activate YOUR-KEY
Golden rule: shadow first. Let shadow mode run through one real work session. When the numbers make sense to you, enable modules one at a time. Everything is reversible with a click or a command.
The first 10 minutes checklist
Step
How to know it worked
harnessup doctor
Hooks installed and events arriving, not just registered.
harnessup open
The panel opens at 127.0.0.1:7717.
harnessup scan
Local history appears with usage per session.
harnessup shadow on
The panel starts showing predicted savings without applying them.
harnessup license status
Current plan and limits are visible.
If you use the API
It works the same: HarnessUp does not proxy or call the API with your credentials. It runs the official claude and codex with your account. With API usage, savings show up directly on the monthly bill — and the panel's counterfactual is even easier to verify.
Understand · 7 min read
How it works
The harness calls a local hook. The daemon normalizes the event, runs modules in series and returns a decision in the format the harness understands. Nothing leaves the machine along the way.
This applies to the whole turn. A deny is not negotiated away by a later allow.
Latency budget and fail open
The budget is a p95 under 50 ms. The hard ceiling is 2 seconds: if the engine does not finish, the hook returns {} and the harness proceeds without a decision. Fail open is mandatory — daemon down, hook fails and the agent keeps working. HarnessUp never holds your session.
Hook measured in production: p95 of ~9 ms. The vault writes in ~9 ms (p99 of 40 ms). The numbers are published in the panel and in the project measurements.
Measurement: token-turns
HarnessUp reads the harness's local transcripts. They carry real token usage per turn, so there is no estimate when a measurement exists. The central unit is the token-turn: tokens multiplied by the turns the content survives in context.
Savings appear in three slices in the savings table: what was cut, what was avoided and what stopped being re-read. The counterfactual — what would have happened without the tool — is always labelled as such. No portfolio number is presented as if it were yours.
Shadow mode
Shadow mode runs the calculation for every module without applying any action. It records applied=0 and returns ActionNone to the harness. It is the conversion engine and the validation of every promise: first you see the savings, then you decide to turn it on.
License, account and relay
Creem is the merchant of record. The daemon does not implement billing: it validates the license and caches it locally. Because Creem routes require the store key, a relay runs on the HarnessUp server and the daemon talks to it without carrying credentials. Without a license, measurement continues; paid modules run in shadow mode.
Where files live
Path
What it stores
~/.harnessup/config.json
Preferences, modules and calibration.
~/.harnessup/license.json
Validated license and cache.
~/.harnessup/runtime.json
Local port and token (0600 permissions).
~/.harnessup/harnessup.db
Measurement, usage and history.
~/.harnessup/graph/<project>.db
The project's context graph.
~/.harnessup/vault/
Originals kept before rewriting, encrypted.
~/.harnessup/quarantine/
Intercepted files, encrypted, 7-day TTL.
~/.claude/settings.json
Claude Code hooks. Merged with backup, never overwritten.
Blocks written into CLAUDE.md and AGENTS.md are delimited by HarnessUp comments. Disabling the module removes the matching block.
Understand · 6 min read
Concepts
The product vocabulary is short and stable. These are the terms that appear in the panel, the CLI and the reports.
Term
Definition
harness
The host agent: Claude Code, Codex, Cursor, OpenCode.
message
From your prompt to the final answer. The Stop event closes it.
turn
Each round trip to the model. One usage line.
call
A tool executed at the end of a turn.
session
Several messages, from opening to shutdown.
token-turn
Tokens multiplied by the turns the content survives in context. The product's central unit.
context debt
The sum of token-turns in a session or period.
module
A toggleable unit with an ID, plan, events and a Handle.
front
One of ten tool groups: metering, monitoring, savings and so on.
shadow mode
Calculate the savings without applying the action.
vault
Where the original is kept before a rewrite. Encrypted and local.
quarantine
Where an intercepted file stays instead of being deleted. 7-day TTL.
cut
The savings slice actually measured in the savings table.
budget
The latency allowance per event and the declarative policy on the Budget & Behavior screen.
panel
The embedded UI served at 127.0.0.1:7717.
context graph
A local graph of decisions, pitfalls, measurements and tasks for the project.
Token-turns in practice
An 8k-token file read on turn 1 of 30 does not cost 8k tokens. It costs 240k token-turns. That is why the tool that saves the most is not the one that writes less: it is the one that removes from context what has already been used.
The panel shows context debt per session and cuts per module. The sum of cuts is what appears as realized savings — kept separate from the counterfactual.
Turn, message, session: do not confuse them
Turn is the smallest billing unit: one model response.
Message starts at your prompt and ends at Stop. It can contain several turns.
Session groups messages and is what Renew Session decides to continue or restart.
Plans in the vocabulary
FREE, Pro, Ultra and GIGA UNLIMITED appear on modules and licenses. A module outside your plan does not disappear: it runs in shadow mode. The agent never breaks because of a plan limit.
Use · 9 min read
Tools
151 tools across ten fronts. Each one is a toggleable module with its own plan, events and behavior — and all of them are reversible.
See where context was spent, conversation by conversation, with the numbers from your own history.
14 toolsLocal panel
Turn it on in the panel. Outside your plan, the tool runs in shadow mode: it measures, applies nothing and deletes nothing.
Deep Scan
Basic FREE
Scans the conversations already saved on the machine and shows where context was spent, one by one.
How it works. The read starts when you trigger the scan in the panel: it walks the local transcripts on demand, counts the sessions it finds and shows the spend broken down by source. It does not change the conversations, and the result card is generated on your computer without sending anything out. On Basic FREE the layer measures and does not apply cuts.
Where. Panel › Metering › Deep Scan
Harness History
Basic FREE
The receipt for every closed conversation: total cost and how the spend breaks down.
How it works. It lists the sessions already closed and opens the receipt for each one, so you can sort by the ones that weighed most without adding everything up by hand or waiting for the invoice. The values come from the local transcripts, not from a fixed table.
Where. Panel › Metering › Harness History
Inside the Wallet
Basic FREE
Splits new input, cache read, cache written and output by model.
How it works. It picks a recorded session and opens the four bands side by side, with the model for each one — the same split the demo uses to show why cache read dominates a long conversation. You see which band weighs on the bill from what your sessions recorded.
Where. Panel › Metering › Inside the Wallet
Cache Cycle
Basic FREE
Shows how much reading the cache cost and when that cost started.
How it works. It points to where in the history a block entered and in how many following rounds it was read again, explaining why a long conversation got expensive. It reads what is already recorded: it measures the cycle without touching the session.
Where. Panel › Metering › Cache Cycle
Openness Meter
Basic FREE
Opens the context of the first turn: what went in, where it came from and how much each integrated MCP server weighs.
How it works. It takes an existing session and breaks down the initial context by source, with the weight of each MCP server. It is how you see the baggage installed before the first question, without opening a new conversation.
Where. Panel › Metering › Openness Meter
Harness Meter
Basic FREE
Measures the core that Claude Code loads before any request and compares its versions.
How it works. It keeps the measurement per harness version and puts one beside the other when you update, so the difference in weight becomes visible. The Harness Update notice uses the same math to say something only when the change passes a minimum.
Where. Panel › Metering › Harness Meter
Insights
Basic FREE
Recalculates real sessions under another rule: a cut, a cap or a context diet.
How it works. You choose the rule and the scope; the panel replays the recorded history and shows the simulated scenario beside the recorded spend. The difference is a hypothesis, not savings already made — the reading only counts after comparing the same base.
Where. Panel › Metering › Insights
Model Insights
Basic FREE
24 questions about cost, speed, errors and reasoning of the models in your use.
How it works. Every answer comes from your history and carries the math behind the number, so you can check the source before switching models. The scores are assessments, not raw token counts.
Where. Panel › Metering › Model Insights
HarnessUp Bench
Basic FREE
Evaluates models on Assertiveness, Optimization and Spend, with a verdict per request.
How it works. It grades each delivery on the three measures and shows who gets it right more often and what each right answer costs. Contributing to the public benchmark is optional and anonymous.
Where. Panel › Metering › HarnessUp Bench
Work Analytics
Basic FREE
Gathers what you ask the agent, what works and what HarnessUp does in its place.
How it works. It groups the requests of the period by task type and crosses each one with the outcome and with the module actions, so you can recognize the request that works. The picture comes from local history, with no invented score.
Where. Panel › Metering › Work Analytics
Bad Use
Basic FREE
Points out 5-hour windows spent with no return and explains why.
How it works. It examines each plan window and flags the ones that yielded little, with the reason — a simple task on an expensive model, a subagent that used too much — and guidance to avoid repeating it. The reading happens in the panel, over the usage already recorded.
Where. Panel › Metering › Bad Use
Plan Usage
Basic FREE
Shows how much each week and each window yielded up to the limit renewal.
How it works. It puts the windows side by side with the credit left over or expired unused, so you can adjust the pace of work. The numbers come from your recorded usage and do not change the plan fee.
Where. Panel › Metering › Plan Usage
/diagnostic
Ultra
Asks for an X-ray of past usage right in the conversation, with the problems and where the spend concentrated.
How it works. You type the command in the chat and get the diagnosis inside the agent itself, without leaving to look for the screen. It is the conversation shortcut for the reading the metering panel already does.
Where. Command /diagnostic
/simulate
Ultra
Tests a change of plan, model or habit against the real history.
How it works. The command runs the hypothetical scenario with your numbers and returns the comparison before you change anything. The difference is a simulation, not money already saved.
Where. Command /simulate
Catalog · 8 tools
Monitoring tools
Follow the spend, the rounds and the subagents while the work moves.
8 toolsLocal panel
Turn it on in the panel. Outside your plan, the tool runs in shadow mode: it measures, applies nothing and deletes nothing.
The Numbers
Basic FREE
Brings together the plan details and each usage window, what is slipping away and your achievements.
How it works. It opens the 5-hour and weekly windows with the remaining credit and what expired unused, a starting point to decide how to pace the work. It is a reading of the plan: on Basic FREE the panel informs, without changing the agent's pace.
Where. Panel › Monitoring › The Numbers
Uptime
Basic FREE
Shows which parts of HarnessUp are running and how many times they stopped.
How it works. It tracks daemon, hooks and panel with status and stop count, so you notice when something went down. The records stay on the machine and feed the panel reading.
Where. Panel › Monitoring › Uptime
Reasoning Check
Basic FREE
Points out how many times reasoning fell short of — or overshot — the size of the task.
How it works. It grades each request by the effort used and highlights light tasks with high effort — or the opposite. Monitoring only signals: it does not change the effort for you.
Where. Panel › Monitoring › Reasoning Check
Subagents Work
Basic FREE
Tracks each subagent's work, the model it ran on and the final result.
How it works. It lists the session's subagents with task, model and outcome, so you can check whether the delivery was worth the spend. The reading updates as the session runs.
Where. Panel › Monitoring › Subagents Work
Harness Update
Basic FREE
On the first session of a new Claude Code version, warns if the initial weight changed.
How it works. When the harness updates, the notice shows the variation in what it loads before any request and only appears if the change passes a minimum. One notice per version, without opening Harness Meter.
Where. Panel › Monitoring › Harness Update
Plan Advisor
Basic FREE
Shows how much you would save with tools from other plans, based on measured usage.
How it works. It compares your recorded usage with what the paid tools would do and delivers at most one notice a day. On Basic FREE it measures and shows; the decision to subscribe remains yours.
Where. Panel › Monitoring › Plan Advisor
Realtime Events
Basic FREE
Shows what each agent tool is doing as it happens.
How it works. It opens the chosen session and lists the calls in order, with time, error and outcome for each; you can pause the refresh and filter by actions or notices. It is the live reading before the per-round summary that Session Radar makes.
Where. Panel › Monitoring › Realtime Events
Session Radar
Basic FREE
Breaks down type, outcome, calls, errors, time and spend per round, with an overall assessment.
How it works. You switch sessions, tap a round and see that round's trace with the spend bands. Inferred scores appear labeled as inference, so they do not pass for direct measurement.
Where. Panel › Monitoring › Session Radar
Catalog · 26 tools
Savings tools
Every re-read, giant log and rambling answer weighs on the next round, and HarnessUp cuts that waste.
26 toolsLocal panel
Turn it on in the panel. Outside your plan, the tool runs in shadow mode: it measures, applies nothing and deletes nothing.
Renew Session
Pro
Finds the turn where continuing costs more than starting over and tells you when the switch pays off.
How it works. It tracks the conversation's cost per round and marks the point where renewal pays off, saying in how many turns the switch breaks even and how many times the renewal goal was met. The point is calculated from the session's own history, not from a fixed average.
Where. Panel › Savings › Renew Session
/up
Pro
Brings Renew Session into the conversation and prepares the next session with the thread of the work.
How it works. You call the command when renewal pays off and it assembles the next session's handoff right in the chat, without losing the thread. It is the conversation shortcut for Renew Session.
Where. Command /up
Smart Compress
Basic FREE
Prepares the compaction summary from local history, without calling the model to summarize everything again.
How it works. It acts when the harness is about to compact: it assembles the summary from what is already recorded on the machine instead of paying a call to re-read the conversation. On Basic FREE it measures in shadow mode; application follows the plan you hold.
Where. Panel › Savings › Smart Compress
Response Diet
Basic FREE
Guides the agent to answer straight, without preamble or repeating what is already in the code.
How it works. It steps in on every answer and cuts the warm-up text, so the result arrives shorter in the conversation itself. On Basic FREE it only measures what it would have cut.
Where. Panel › Savings › Response Diet
Code Finder
Pro
On the first read of a large file, delivers function names and line ranges.
How it works. It acts in the read call: returns the file's skeleton and sends the body only when the agent asks, including reads through the terminal. The file stays whole on disk.
Where. Panel › Savings › Code Finder
Smart Cut
Pro
Delivers the first page of large files, reports the total size and teaches how to ask for the rest.
How it works. It intercepts the read of long logs and data and swaps the whole content for the first page, with the size and the path to the remainder. Nothing is deleted: the original file stays on disk.
Where. Panel › Savings › Smart Cut
Delta Read
Pro
Keeps track of which lines the agent has already seen and avoids re-reading what is still the same.
How it works. It compares the new read with what is already in context and refuses or shortens the repeated stretch, warning when the file has not changed. What changed goes in normally.
Where. Panel › Savings › Delta Read
Head Limit
Pro
Caps unlimited content search and teaches the agent to ask for the rest.
How it works. It acts in Grep: cuts the giant result and shows how to ask for the next band; filename searches such as Glob are left out. The limit applies per call.
Where. Panel › Savings › Head Limit
WebFetch Diet
Pro
Asks for answers of up to 700 words when fetching pages and refuses to fetch the same page again in the session.
How it works. It acts in the WebFetch call: asks for a short summary and, if the page was already read in that session, refuses the repeat. Requests for full text or a defined size are honored.
Where. Panel › Savings › WebFetch Diet
Terminal Guide
Pro
Summarizes long test, build and log output, preserving the beginning, the end and the error lines.
How it works. It acts on terminal output: delivers the useful excerpt and keeps the full output on the computer, with a pointer for lookup. The agent sees the error even when the log has hundreds of lines.
Where. Panel › Savings › Terminal Guide
Output Diet
Pro
Limits build, test and search outputs to the beginning and the end of the result.
How it works. It replaces the whole text with the beginning, the end and a pointer to the content saved on disk. It applies to tool output that blows past the cap.
Where. Panel › Savings › Output Diet
Conduction Write
Pro
Blocks a write that would leave the file exactly the same.
How it works. It acts in the write call: compares with the current content and refuses the edit that changes nothing, including one that swaps a text for itself. It saves a call without altering the file.
Where. Panel › Savings › Conduction Write
No Comment
Pro
Asks the agent not to comment the obvious and measures how much new code still arrives commented.
How it works. The guidance goes into the request and the measurement stays in the panel, so you can see whether it is being followed. It does not delete existing comments: it guides what comes next.
Where. Panel › Savings › No Comment
First Pass
Pro
Compiles the project on the machine after an edit and shows the error to the agent before the next turn.
How it works. It fires after the edit: runs the build and, if it breaks, the error goes into the next round so the fix goes out with it. The agent does not spend a turn just to find out the code does not compile.
Where. Panel › Savings › First Pass
Way Out
Pro
When a tool fails, points to a fix path that already exists in the project.
How it works. It acts on failure, looks in the project for an example that solves the same problem and cites the source file. The hint only appears when there is evidence.
Where. Panel › Savings › Way Out
Read Router
Pro
Delivers the section index of a large context file first.
How it works. It acts on read: returns the index and waits for the request; if the agent insists, it gets the full text. Nothing is deleted.
Where. Panel › Savings › Read Router
Hot Context
Pro
Slices old history into immutable fragments stored in git and findable by search.
How it works. When the history grows, the old stretch becomes a fragment and the model starts reading a short, current view, with search for the detail. The material remains recoverable in the repository.
Where. Panel › Savings › Hot Context
Context Graph
Pro
Builds a local graph, with no LLM or embeddings, holding the session queue, decisions and measurements.
How it works. At session opening, it delivers the summary with file and line for each item and flags missing context and duplicate decisions. The graph is built on the machine from the project files.
Where. Panel › Savings › Context Graph
Session Protocol
Pro
Writes into the master file the guidance to open context documents only when the opening summary is not enough.
How it works. The rule goes into the master file once and applies to the following sessions, avoiding re-reading everything on each opening. You turn it on in the panel and the text appears in the project file.
Where. Panel › Savings › Session Protocol
Memory State
Pro
Organizes the instructions the agent looks for, teaches it to read only the necessary stretch and caps the size of the files it rewrites.
How it works. Three buttons in the panel touch those three fronts at once, and the change applies to the project's next sessions. It is a file adjustment, not a conversation one.
Where. Panel › Savings › Memory State
Stature Brake
Pro
Caps the context file and archives the excess in SafeBox, recoverable in one click.
How it works. It acts when the context file passes the cap: cuts the excess into SafeBox and points out leftovers scattered through the project. What left comes back in one click.
Where. Panel › Savings › Stature Brake
Rules Auditor
Pro
Audits CLAUDE.md and the rules reloaded on every round.
How it works. It reads the rules that enter every conversation and proposes removing what does not serve, showing the fixed cost of each block. The decision to cut is yours: it audits and suggests.
Where. Panel › Savings › Rules Auditor
Optimize Harness
Pro
Shows the measured cost of skills, subagents, plugins and servers, with switches to turn them on or off.
How it works. Each item appears with the cost it adds to context and a toggle; what was not measured gets no number. It is how you take out of the opening what weighs without being used.
Where. Panel › Savings › Optimize Harness
Cheap Swarm
Ultra
Delegates 26 task classes to subagents with the appropriate model and effort.
How it works. You turn it on in the panel and the cast takes the field according to the task class; the count shows what was delegated, what came back and what the main context stopped paying. The main conversation decides and gathers the results.
Where. Panel › Savings › Cheap Swarm
/cheap
Ultra
Triggers the Cheap Swarm cast inside the conversation.
How it works. Called in the chat, it passes each step a subagent can execute to the cast and returns the results to the main thread. It is the same Cheap Swarm catalog, without leaving the conversation.
Where. Command /cheap
/print
Pro
Brings the most recent screenshots from the Windows screenshots folder into the conversation.
How it works. The command reads the captures folder and attaches the images; with no number, the last three arrive, and you can choose the count or the screen. Nothing is sent off the machine.
Where. Command /print
Catalog · 6 tools
Customize tools
Keep status in sight, dress the app in another theme and get a nudge when it is your turn.
6 toolsLocal panel
Turn it on in the panel. Outside your plan, the tool runs in shadow mode: it measures, applies nothing and deletes nothing.
Info Pills
Basic FREE
One line per session, over the windows: count, status, idle time and tag.
How it works. The pills sit over open windows and let clicks pass through, so you know which session finished without switching screens. What appears in each pill is chosen in the customization bench.
Where. Panel › Customize › Info Pills
Permission Push
Basic FREE
Sends a Windows notification when the agent stops to ask for permission.
How it works. It fires the moment the session starts waiting for an answer and arrives even with the panel closed. The decision to allow stays with you, in the session.
Where. Panel › Customize › Permission Push
Status Bar
Basic FREE
Shows the session receipt in the Claude Code CLI footer.
How it works. Colored pills track id, turns, tokens, US$, idle time and the 5-hour window in real time while the agent works in the terminal. Read-only: it does not change the command.
Where. Panel › Customize › Status Bar
Cache Alarm
Basic FREE
Warns that an idle session is close to cooling the cache.
How it works. The alarm rings in the corner of the screen at 28 minutes of inactivity in Codex and 43 in Claude Code, before the discount drops. Typing again within the window keeps the cache warm.
Where. Panel › Customize › Cache Alarm
Mod Themes
Basic FREE
Installs one of seven themes in the Claude Code desktop app.
How it works. Colored borders separate your messages, the agent's answers and the tools; you switch the theme in the panel and the appearance changes right away. The agent's work stays the same.
Where. Panel › Customize › Mod Themes
Finish Line Push
Pro
Notifies with sound and animation when the agent finishes the task, with confetti, fireworks or a neon frame.
How it works. You choose the effect and whether the sound stays on; at the end of the task, the notice appears even if you are in another window. It is a completion signal: it changes neither the agent's work nor any permission.
Where. Panel › Customize › Finish Line Push
Catalog · 8 tools
Security tools
A brake on the dangerous command, the hidden instruction and the edit that went too far.
8 toolsLocal panel
Turn it on in the panel. Outside your plan, the tool runs in shadow mode: it measures, applies nothing and deletes nothing.
Only Read
Pro
Lets the agent read, explain and propose, but refuses file-editing tools.
How it works. The session stays open in read mode: the agent can look things up and answer, and any attempt to write a file is refused. Shell commands are not blocked by this tool, so pair it with No Touch, No Delete or SEC Station depending on the work.
Where. Panel › Security › Only Read
SafeBox
Pro
Keeps for seven days what HarnessUp intercepted, cut or blocked, so you can check and recover it.
How it works. In the panel it opens the list of material that left the agent's path; each item can be inspected and recovered within the period. It is the layer's interception archive — for project restore points, the right front is Protection.
Where. Panel › Security › SafeBox
No Touch
Ultra
Blocks a folder or a file completely, with no read or write, or write only.
How it works. The limit applies per session or per project and reaches even files inside the project itself; in the write-only option, the agent still reads the content. You choose the path in the panel and it stays untouchable until you change it.
Where. Panel › Security › No Touch
No Delete
Ultra
Refuses deletion commands such as rm, del, Remove-Item, git clean and git rm.
How it works. It inspects shell commands and refuses deletion, with scope per session, per project or for everything. git rm --cached passes, because it removes from the index without deleting the local file.
Where. Panel › Security › No Delete
Smart Security
GIGA UNLIMITED
Brings together the four protections that watch content, secrets, code and personal data: No Injection, No Leak, CodeSafe and SafeCorp.
How it works. Each guard acts at one point: No Injection blocks hidden instructions in consulted content; No Leak follows the sequence of suspicious steps and masks secrets; CodeSafe keeps every code finding open until the fix; SafeCorp blocks personal data in output, with LGPD, GDPR, CCPA, HIPAA and PCI rules. Each guard has its own switch in the panel.
Where. Panel › Security › Smart Security
Sentinel
GIGA UNLIMITED
Reads every write and every command from the agent in real time and blocks the action when it identifies a security flaw.
How it works. It acts before the write: if the write or the command opens a flaw, the action is refused and the panel points out what to fix. The fix remains the agent's work.
Where. Panel › Security › Sentinel
SEC Station
GIGA UNLIMITED
Locks groups of dangerous commands across all sessions.
How it works. Each group — curl, sudo, writing to the shell profile, package publishing — has its own switch in the panel, and the setting applies to every session. You choose what stays blocked.
Where. Panel › Security › SEC Station
Security Check
GIGA UNLIMITED
Reviews the project in six layers and assembles a preview of a PDF report.
How it works. It walks secrets in the project, MCP/hooks/permissions, hidden instructions, supply chain and CI, containers and code, and dependencies. The result stays in the panel for reading before it becomes a PDF, and points out what to review in each layer.
Where. Panel › Security › Security Check
Catalog · 10 tools
Protection tools
Damage can be undone and the secret never even enters the conversation.
10 toolsLocal panel
Turn it on in the panel. Outside your plan, the tool runs in shadow mode: it measures, applies nothing and deletes nothing.
Smart Return
Pro
Undoes a turn's file changes, from the command you choose.
How it works. It only acts with the session stopped: you mark the turn's step and click undo, and the files return to the previous state. Changes made by shell commands are left out.
Where. Panel › Protection › Smart Return
Checkpoint
Pro
Keeps a project restore point inside git, under its own reference.
How it works. You mark it before a big change and the state is saved without touching the files, what is staged for commit or your branches. If the change goes wrong, the point is still there waiting.
Where. Panel › Protection › Checkpoint
Git Shield
Pro
Checks the project .gitignore, creates one if there is none and writes the missing lines.
How it works. It acts in the project: adds a marked block with secrets, generated folders and tool leftovers. One click removes the whole block; files, staging and branches are not touched.
Where. Panel › Protection › Git Shield
Eyes Off
Pro
Takes screen token spend away from the model and leaves the visual check to you.
How it works. It blocks browser, screenshot and remote control, and the answer ends with a list of what you should check. The checklist stays in the panel.
Where. Panel › Protection › Eyes Off
Secret Lock
Ultra
Swaps secret values for a mask before they enter the conversation.
How it works. It acts on read — the agent sees variable names with masked values — and on the request, holding back a password pasted by accident; the file on disk stays the same. It comes on by default on Ultra.
Where. Panel › Protection › Secret Lock
Deny MCP
Ultra
Turns off MCP tools for one session or for all of them.
How it works. The server you do not trust leaves the path until you turn it back on, and the work continues without it. The door stays closed in the panel, per session or for all.
Where. Panel › Protection › Deny MCP
Blackout
Ultra
Takes the session off the network.
How it works. It refuses WebFetch, WebSearch, curl, git push and npm install in the chosen session or in all of them, until you turn it back on. The rest of the computer stays normal: it is the agent that goes without network.
Where. Panel › Protection › Blackout
Sandbox Mode
Ultra
Locks the session into a folder.
How it works. Reading stays free, but deleting and overwriting outside the folder are blocked — useful for testing a new agent. The limit applies to the chosen session.
Where. Panel › Protection › Sandbox Mode
UltraShield Guard
Ultra
Gives a wrong command a safety net, with three reactions according to the size of the risk.
How it works. Deletion outside the project is blocked; what is deleted inside the project is kept so it can come back; a risky git command gets a copy beforehand and still asks for your confirmation. Each reaction acts at the moment of the command.
Where. Panel › Protection › UltraShield Guard
Antiloop Protector
Ultra
Warns the agent that it is repeating itself and stops the session when spend and repetition move together.
How it works. The warning comes first and the agent can change route; the stop only happens when spend goes well above the project usual level and the same actions repeat without progress. There are two switches: you can turn off the spend brake and keep the warning.
Where. Panel › Protection › Antiloop Protector
Catalog · 5 tools
Builders tools
You tell the idea and HarnessUp interviews, writes the PRD and the SPEC, raises the project and builds the site.
5 toolsLocal panel
Turn it on in the panel. Outside your plan, the tool runs in shadow mode: it measures, applies nothing and deletes nothing.
SPEC / PRD Express
GIGA UNLIMITED
Researches the competition, interviews you, criticizes weak paths and writes the PRD and the SPEC.
How it works. The conversation starts from the idea and moves into an interview, with criticism of what looks weak before it becomes code. At the end it proposes three stack plans — Sprint Method, Balanced Mode and Triple-A — and, with the chosen one, writes PRD.md and SPEC.md into the project folder.
Where. Panel › Builders › SPEC / PRD Express
DOCs Generator
GIGA UNLIMITED
Generates the project documentation in a folder named after it.
How it works. The folder comes out in the format /bones reads, so the project does not start without written documentation. You run it from the panel and the files stay in the project itself.
Where. Panel › Builders › DOCs Generator
/bones
GIGA UNLIMITED
Reads the documentation in the folder and raises the API contract, the data model, minimal backend and frontend and an end-to-end slice.
How it works. It works in waves and phases, without asking questions: when it finishes, there is a slice working end to end. The real order of the phases is decided by what the documentation asks for.
Where. Command /bones
DOC for Website
GIGA UNLIMITED
Scans the project, inventories the features that can be sold and writes the product dossier.
How it works. It interviews you for up to 8 rounds before writing, so the text is born from the code without promises the product does not keep. The dossier stays in the folder and serves as the base for Website Builder.
Where. Panel › Builders › DOC for Website
Website Builder
GIGA UNLIMITED
Interviews you, proposes style and sections and builds a static site in your folder.
How it works. The preview appears in the panel during the build, which runs with a cost cap set by you. It requires no programming knowledge, and the final site stays in the project folder.
Where. Panel › Builders › Website Builder
Catalog · 25 tools
Control tools
Spend cap, working hours, rules that do not evaporate mid-conversation and gates before ending the session.
25 toolsLocal panel
Turn it on in the panel. Outside your plan, the tool runs in shadow mode: it measures, applies nothing and deletes nothing.
Force Test
Pro
Keeps the session from ending when the code changed and no test passed after the change.
How it works. The agent goes back to fix it and the session only ends with a green test; after 3 blocks without a test, the session ends, so the brake does not become a prison. Each gate is a separate switch in the panel.
Where. Panel › Control › Force Test
Enforce TDD
Ultra
Only releases code file editing after a test written or run red since the last request.
How it works. Documentation and configuration pass straight through, and the TDD rule goes into every request. It works as a phase gate at the end of the turn.
Where. Panel › Control › Enforce TDD
Force Lint
Ultra
Requires the project linter to pass before the changed code goes out.
How it works. The agent runs it, fixes what the linter points out and is forbidden from loosening the configuration to close the task. The session does not end with lint broken.
Where. Panel › Control › Force Lint
Doc Now
Ultra
Forces the project documentation to be updated before the turn of whoever edited code ends.
How it works. TODO.md and DECISIONS.md are part of the count, and the agent goes back to write what is missing. After 3 blocks in the same turn, it gives up.
Where. Panel › Control › Doc Now
Budget Cap
Pro
Sets spend and usage guidance per task, session or project and tells the model how to work within them.
How it works. It acts in the request as an agreement with the model, with no automatic cut; when the limit is exceeded, nothing is interrupted. For the actual cut, Stop Loss is the one that acts.
Where. Panel › Control › Budget Cap
Stop Loss
Pro
Refuses the next tool call when the session passes the cap in US$, tokens or turns.
How it works. You choose the measure and the limit in the panel; when it goes over, the next call is refused and everything closes in order, with what was in progress finishing first. It is the brake that actually stops the session.
Where. Panel › Control › Stop Loss
Do Not Reply
Pro
Removes the button to dismiss Claude Code's next-step suggestion.
How it works. Without the click, the model does not re-read the whole context just to hear that nothing more is needed. The change is in the interface and applies to the session where it is on.
Where. Panel › Control › Do Not Reply
Rule Checker
Pro
At the end of each answer, checks the ToolBox rules that leave a trace.
How it works. It looks for a cited file that was never opened, a loosened test and an escape hatch on types; if it finds a failure, it notifies you in the panel. It does not block anything.
Where. Panel › Control › Rule Checker
Init Convention
Pro
Delivers a text or up to 6 files once, at the opening of each new session.
How it works. It applies globally or per project; a resumed, compacted or branched session does not get it again. It is the convention that does not need repeating every conversation.
Where. Panel › Control › Init Convention
Reasoning Advisor
Pro
Suggests the model and the effort the task calls for.
How it works. The model comes in the first prompt, when switching still does not invalidate the cache; after that, only the effort. The suggestion appears at the start of the work and you decide whether to accept it.
Where. Panel › Control › Reasoning Advisor
Attach Skill
Pro
Attaches up to 8 of your skills to a moment of the session and puts them before the model when it arrives.
How it works. The catalog gathers skills from the user, from plugins and from open projects, with the description in view. The skill enters at the chosen moment, not at the opening of everything.
Where. Panel › Control › Attach Skill
Attach Files
Pro
Attaches key project files to a moment of the session and points the model to them at the right time.
How it works. You define the file and the moment, and the material enters context when the time comes — which helps on a short task inside a big project. You do not have to explain again where each thing is.
Where. Panel › Control › Attach Files
Working Hours
Pro
Defines when agents may work.
How it works. Outside the hours, every tool is refused in any session open on the computer. Each day accepts up to 4 intervals — morning, afternoon, a breather at lunch and an evening window.
Where. Panel › Control › Working Hours
Scheduled Break
Pro
Schedules breaks with a countdown on screen.
How it works. During the break, tools are refused and new requests wait until the time runs out. The session stays open and resumes by itself at the end of the countdown.
Where. Panel › Control › Scheduled Break
Git Admin
Pro
Locks the most sensitive git actions across all sessions.
How it works. Commit and push to main, force push, history rewrite, publishing and deploy each have their own switch, so you release one without opening the others. The lock applies to every session.
Where. Panel › Control › Git Admin
No Subagents
Pro
Refuses the opening of subagents.
How it works. The call is refused and the agent does the work in the conversation itself, with no unexpected delegation on the bill. You turn it on when you need work without branching.
Where. Panel › Control › No Subagents
/watch
Ultra
Puts an agent watching the other open sessions, without interfering with their work.
How it works. The observer follows what the other sessions do and, at the end, delivers an opinion on what it saw. It does not touch the observed work.
Where. Command /watch
Swarm Commander
Ultra
When the main model would do delegable work by itself, sends the task to the Cheap Swarm cast.
How it works. It watches the session in real time and routes the task to the appropriate subagent. It only acts with Cheap Swarm on.
Where. Panel › Control › Swarm Commander
Productivity Mode
Ultra
Holds a one-line request and reminds you to gather 3 or 4 into one message.
How it works. If the same request is sent again within 30 minutes, it goes straight through, so it does not jam someone who changed their mind. The idea is to cut the round that did not need to exist.
Where. Panel › Control › Productivity Mode
Stick Prompt
Ultra
Repeats the rule the agent keeps forgetting on every prompt of yours, until you take it off.
How it works. The rule comes back in every request, inside the context, without you rewriting it. You turn it on and off in the panel, per project.
Where. Panel › Control › Stick Prompt
Anti-fading
Ultra
Sends a silent reminder every N turns, from 1 to 100.
How it works. The reminder enters context at the chosen interval, with up to 1,000 characters, and does not appear in the conversation. It applies to one session, one project or all.
Where. Panel › Control › Anti-fading
Subs Director
Ultra
Gives each subagent a budget of turns and time.
How it works. It warns at 50% of the budget, corrects course at 75% when the subagent starts reading outside the task and calls for the report at 90%. No one blows the limit or comes back empty-handed.
Where. Panel › Control › Subs Director
Anti-Collision
GIGA UNLIMITED
Two sessions on the same file or folder? The agent on the larger model wins.
How it works. The smaller one waits idle, without running a turn or burning window, until its turn is released. It applies to several sessions in the same project.
Where. Panel › Control › Anti-Collision
ToolBox
Pro
Gathers 17 ready-made conventions for the agent to work with discipline.
How it works. Each convention switches on by itself and shows the cost per turn in view — diff size, one thing at a time, a test that fails. The rules enter the prompt at the right time, not all at once.
Where. Panel › Control › ToolBox
HarnessUp OFF
Basic FREE
Turns off everything HarnessUp does for a while, so you can work with the bare agent.
How it works. You choose the period and, at the end of it, everything comes back by itself, without uninstalling anything. It is useful to compare with and without the layer or to debug a problem.
Where. Panel › Control › HarnessUp OFF
Catalog · 26 tools
Productivity tools
Build, commit, test and deploy leave the model's path, and the rest runs by itself in the panel.
26 toolsLocal panel
Turn it on in the panel. Outside your plan, the tool runs in shadow mode: it measures, applies nothing and deletes nothing.
HarnessUp Flow
GIGA UNLIMITED
Creates flows from the chat, in natural language, on an infinite canvas.
How it works. Each point has a tool and a model, and the links have conditions written the way you speak; there are MCP, API, A2A, webhook and browser connectors, login to 16 services and proxy, cloud browser and Android per point. You drag the points and the links follow.
Where. Panel › Productivity › HarnessUp Flow
HarnessUp Board
GIGA UNLIMITED
Infinite board with notes, text, shapes, arrows, images, videos, files, link previews and Kanban columns.
How it works. You build the board by dragging, and the backup goes to the synced folder you choose — Drive, OneDrive, Dropbox or iCloud. The same board feeds /board.
Where. Panel › Productivity › HarnessUp Board
/board
GIGA UNLIMITED
Type /board in the chat and the whole board becomes context for the model.
How it works. Notes, arrows and columns arrive as they are, without you rewriting the plan in text. The command reads the Board of the open project.
Where. Command /board
Build Now
Ultra
Builds the project at the end of the turn in which files changed, without spending a model turn.
How it works. If the build fails, the error goes back to the agent and the fix continues in the normal work; there is also a Build Now button in the panel. The build runs outside the conversation.
Where. Panel › Productivity › Build Now
Smart Commit
Ultra
Makes the commit at the end of the task, without an extra round of conversation.
How it works. It does not push; each commit goes into a list in the panel with the option to undo. Pushing to GitHub is left to GitHub in 1 Click, which asks for confirmation.
Where. Panel › Productivity › Smart Commit
Merge Easy
Ultra
Runs the merge end to end, with conflicts resolved by the model and then checked.
How it works. It runs behind a backup and with a cost cap for the resolution; if something fails, the merge is aborted and you return to the source branch. You choose source and target in the panel.
Where. Panel › Productivity › Merge Easy
GitHub in 1 Click
Ultra
Creates a private repository and pushes the code with the official gh, already logged in by you.
How it works. The push only happens after confirmation in the panel, so nothing goes up without you seeing it. The account used is the one you already have on the machine.
Where. Panel › Productivity › GitHub
Deploy in 1 Click
Ultra
Publishes with the official tool of each hosting service and the login you already have on the machine.
How it works. Each service follows its own command, chosen in the panel, with no new credential to paste. The deploy uses the account you already use.
Where. Panel › Productivity › Deploy
Test Run
GIGA UNLIMITED
Runs the project tests from the panel, without spending an agent turn and without opening the terminal.
How it works. You trigger it on the spot or schedule it to run by itself at the times you choose. The result appears in the panel.
Where. Panel › Productivity › Test Run
Test Now
GIGA UNLIMITED
Finds untested code and weak, skipped or broken tests in a list that is easy to read.
How it works. You mark the items and the chosen session writes and runs the tests. Good for a project that grew fast and was left with uncovered spots.
Where. Panel › Productivity › Test Now
Lint, Format, Compile
GIGA UNLIMITED
Gathers lint, formatting and compilation in one place, with the right command for each language in the project.
How it works. The panel picks the command for each stage, so you memorize none, and shows the result of each one. Everything runs in the right folder.
Where. Panel › Productivity › Lint, Format, Compile
Command Control
GIGA UNLIMITED
Runs project commands or your own commands straight from the panel, already in the right folder.
How it works. You schedule everyday routines to run by themselves, without opening a terminal. The commands stay saved in the panel.
Where. Panel › Productivity › Command Control
Super Code Review
GIGA UNLIMITED
Assesses the code and shows the composition, a score from 0 to 100, the recommended stack and a better architecture.
How it works. It points out bugs and security flaws in a complete PDF report. The review starts from the project that is in the folder.
Where. Panel › Productivity › Super Code Review
Real Time Advisor
GIGA UNLIMITED
Advises the agent while it writes on the modern, simple and robust way to build each part.
How it works. It also warns about junk in the project and forgotten documentation, at the moment the agent works. The hint enters the session context.
Where. Panel › Productivity › Real Time Advisor
Recombine
GIGA UNLIMITED
Looks for duplication and verbosity and proposes slimming the code down.
How it works. The session refactors on a separate branch, runs tests and build and asks before taking it to main. Nothing goes to main without your answer.
Where. Panel › Productivity › Recombine
Deflate
GIGA UNLIMITED
Finds unused code, files and dependencies, debugging leftovers and committed artifacts.
How it works. Nothing is deleted on its own: each item appears in a list and you decide its fate. The same goes for debugging leftovers.
Where. Panel › Productivity › Deflate
/analyse
Ultra
Reviews the last commits in depth, straight from the chat, with strengths, weaknesses, gaps and open holes.
How it works. You call the command and get the review in the conversation, without opening another tool. It serves those who work alone and have no reviewer looking at every commit.
Where. Command /analyse
/save
Ultra
Saves the current conversation as Markdown right away.
How it works. It works in the middle of a long task, without waiting for the automatic save and without calling the model. The file comes out faithful to what was said.
Where. Command /save
Smart Report
Ultra
Generates the account of what HarnessUp measured and did, in PDF, Markdown or text.
How it works. Every number comes with its source, so you can show it to someone or check it yourself. The export is triggered in the panel.
Where. Panel › Productivity › Smart Report
Notify-me
Ultra
Puts your message into the conversation with the agent at the moment of the work you choose.
How it works. You write the note, choose the moment and can include markdown and attachments. The message goes in by itself, without you waiting for the time.
Where. Panel › Productivity › Notify-me
Actions
Ultra
Sends a message and files to a webhook, an API or an MCP server at a chosen moment of the session.
How it works. The trigger happens at the set time, with no copy and paste between tools. You configure the destination in the panel.
Where. Panel › Productivity › Actions
Ez Work
Basic FREE
Gathers everyday shortcuts in the panel, such as stopping the right session and opening the next one already with the handoff.
How it works. Each tool in the set has its own plan and acts in the panel, without spending a context round. It is the workbench the others come from.
Where. Panel › Productivity › Ez Work
Context Central
Ultra
Keeps a short record per session and six drawers of durable context.
How it works. The record points to where the detail is instead of copying it, and all the Code Deck slots of a project share the same notebook. The next session starts knowing where it stopped.
Where. Panel › Productivity › Context Central
Code Deck
GIGA UNLIMITED
Runs more than 17 agent CLIs in up to 4 slots at the same time, in one window.
How it works. Each slot has its own terminal and chat mode; the Synergy mode leaves the reviewing session waiting without spending credit. The automatic integration works today for Claude Code and Codex, and the sessions also appear in the Session Panel.
Where. Panel › Productivity › Code Deck
KanBan Sync
GIGA UNLIMITED
Builds the cards from what the agent plans and from the project queue.
How it works. Each card is born from the TODO, the state and the agent's plan, and shows the cost actually measured; the column follows the progress. You follow the project without loose notes.
Where. Panel › Productivity › KanBan Sync
Save my Session
Ultra
Saves conversations as Markdown automatically, without calling the model.
How it works. You choose the format — full, summarized or dialogue only — and the file comes out faithful to what was said. The saving happens throughout the work.
Where. Panel › Productivity › Save my Session
Catalog · 23 tools
Utilities tools
Know the price of the prompt before sending, stop the agent without paying for a round and find the right session among several open.
23 toolsLocal panel
Turn it on in the panel. Outside your plan, the tool runs in shadow mode: it measures, applies nothing and deletes nothing.
Token Calculator
Basic FREE
Turns a text or a token count into cost, at each billing band of the model.
How it works. You paste the prompt and see the price before sending, with the spend measured in your sessions beside it. It is how you forecast the cost of a large prompt without doing the math by hand.
Where. Panel › Utilities › Token Calculator
Frozen Cache
Basic FREE
Warns when an idle conversation is close to losing the cache discount.
How it works. The warning comes as a system notification; coming back and typing while the conversation is alive keeps the discount for free. Nothing in the session is changed.
Where. Panel › Utilities › Frozen Cache
Cache Lock
Basic FREE
Warns before an action that would drop the discount, such as switching model or plugin in the middle of the work.
How it works. It also holds the return to a conversation idle for more than 55 minutes, so coming back does not cost the whole context. You decide whether to go ahead.
Where. Panel › Utilities › Cache Lock
Caching Counter
Basic FREE
When the discount drops, says why.
How it works. It covers changes at the start of the conversation, such as editing CLAUDE.md with the session open, so the same drop does not repeat. The reading stays in the panel.
Where. Panel › Utilities › Caching Counter
BTW Notice
Basic FREE
Notices the side question and suggests asking it with /btw.
How it works. /btw answers without growing the conversation, so context does not swell with what is not part of the task. The suggestion appears at the moment the side question comes up.
Where. Panel › Utilities › BTW Notice
Prompt Optimizer
Basic FREE
Rewrites the prompt shorter and more direct, cutting courtesy, detour and repeated phrasing.
How it works. Local rules make the first cut and the official claude on your account tries a smaller version; code between backticks stays intact. If no version comes out smaller, the original returns unchanged.
Where. Panel › Utilities › Prompt Optimizer
MetaPrompt · Prompt Creator
Basic FREE
Turns an idea described in your own words into a complete prompt, ready for another AI tool.
How it works. There are 34 types in 7 groups; you choose the type, the language and the extra details, and the history keeps the last 30 creations. The output comes with title, prompt, parameters and variations.
Where. Panel › Utilities › MetaPrompt
Easy Prompt
Basic FREE
Stores the prompts you repeat every day.
How it works. One click copies and you just paste it into the conversation with the agent, without rewriting the request. The list stays in the panel.
Where. Panel › Utilities › Easy Prompt
Ghost Prompt
Ultra
Places a phrase of yours inside the running session, at the next tool use, without appearing as a chat message.
How it works. The phrase arrives in the next call, so the route is corrected without spending a whole round. You choose the session in the panel.
Where. Panel › Utilities › Ghost Prompt
Mirror Master
Pro
Copies the master instructions file of one harness into the format of the others.
How it works. The conversion is done with no model and no cost, and the same rules start applying in CLAUDE.md, AGENTS.md, GEMINI.md and the other supported formats. You choose the source and the target.
Where. Panel › Utilities › Mirror Master
Session Panel
Ultra
Brings together on one screen the sessions of every account and origin: CLI, desktop app, IDE, SDK and Code Deck.
How it works. Each session becomes a card with branch, worktree, account, model, rounds and cost, and there are filters by project, flagged and important. This is where the stop, pause and open-next actions come from.
Where. Panel › Utilities › Session Panel
Tag My Session
Pro
Gives each open session a short tag and a note.
How it works. Both also appear in the Session Panel, so you know which is which at a glance. The tag follows the session while it is open.
Where. Panel › Utilities › Tag My Session
Work Timer
Basic FREE
Shows two timers in the corner of the screen: the time the agents worked and yours.
How it works. It accepts up to 5 time zones and keeps the history per day, so you can compare the work of each day. The clock runs in the corner, without taking up the conversation.
Where. Panel › Utilities › Work Timer
Stop Agent
Basic FREE
Stops the chosen session at the next tool call, with no message in the chat.
How it works. The order arrives in the next call, so the model does not re-read the whole context just to receive the stop. You choose the session in the panel.
Where. Panel › Utilities › Stop Agent
Coffee Brake
Pro
Pauses the session without ending it.
How it works. The agent waits without spending credits until you resume with one click, and continues from where it stopped. The session stays open during the pause.
Where. Panel › Utilities › Coffee Brake
Port Guard
Basic FREE
Forbids the agent to kill a process or port.
How it works. So it does not bring down the local server you left running; what it put up itself, it can still stop. The lock acts on the shell command.
Where. Panel › Utilities › Port Guard
Copy Handoff
Pro
Joins the most recent handoff from Renew Session and from Context Graph into one image.
How it works. In the new session, one Ctrl+V brings the context along, without pasting text by text. The image is generated on your machine.
Where. Panel › Utilities › Copy Handoff
Trash Killer
Pro
Looks for work leftovers scattered through the project — .tmp, .bak, logs and drafts — and puts everything in quarantine.
How it works. Nothing is deleted for good: one click restores what went to quarantine. You see the list before acting.
Where. Panel › Utilities › Trash Killer
Temps Clean
Pro
Cleans the residue agents leave on disk.
How it works. Before deleting, it shows the count and size of what leaves and asks for your confirmation — it is the only HarnessUp action that deletes without keeping a copy. Nothing leaves without your yes.
Where. Panel › Utilities › Temps Clean
Fragment Generator
Basic FREE
Takes pasted text or a screenshot, identifies what it is and saves the stretch in the right folder, with a name.
How it works. It separates code, data, log, prompt or note, and the screenshot becomes text through the OCR built into Windows, without leaving the machine. The file is saved already named.
Where. Panel › Utilities › Fragment Generator
HarnessUp Shop
Basic FREE
Curated catalog of inspected agents, skills, MCPs and tools.
How it works. Installation is one click through the Safe Installer, instead of downloading from anywhere. Only what passed inspection enters the catalog.
Where. Panel › Utilities › HarnessUp Shop
/explain
Pro
Re-explains in plain language, for people outside the field, the stretch you did not understand.
How it works. You call the command on the answer that is too technical and get the explanation in the conversation. It serves mid-work, without opening another tool.
Where. Command /explain
/ask
Pro
Explains the agent's last decision question: what it suggests, the pros and cons of each option and what changes for whoever uses the product.
How it works. The choice stays with you, and it does not implement before your answer. The command summarizes the decision that was left open.
Where. Command /ask
Use · 7 min read
Panel and CLI
The panel handles the everyday work without a terminal. The CLI handles what must be scriptable — and both talk to the same local daemon.
The panel
Run harnessup open (or harnessup panel) and go to 127.0.0.1:7717. The UI is embedded in the binary, with more than 200 toggles: turn on, turn off, calibrate and undo everything in one click. Nothing needs admin and nothing is sent to the cloud.
Measurement: usage per session, model and period; token-turns and context debt.
Savings: cuts per module, shadow mode and simulation before applying.
Control: Budget & Behavior, Stop Loss, schedules and rules.
Security: vault, quarantine, undo and interception history.
License: current plan, limits and activation.
Most used commands
Command
What for
harnessup doctor
Hooks, events and channel diagnostics.
harnessup status
Daemon and installation status.
harnessup open
Opens the local panel.
harnessup scan
Reads local history and measures what already happened.
harnessup report
Usage and savings report.
harnessup simulate
Counterfactual over real history.
harnessup forecast
What has been spent and where it is going.
harnessup guard
Spend brake and blocking policy.
harnessup shadow on|off|status
Controls shadow mode.
harnessup modules list|enable|disable|shadow
Manages modules individually.
harnessup license status|activate|deactivate
Plugs and unplugs the license.
harnessup decision-graph search <term>
Queries the project's decision graph.
harnessup context
Opens the project context and handoff.
harnessup undo
Reverts the last turn.
harnessup vault
Lists what is kept in the vault.
harnessup quarantine
Lists and restores quarantined files.
harnessup restart
Restarts the daemon.
harnessup version
Running version.
There is no harnessup budget. Budgeting is harnessup guard (the brake) and harnessup forecast (what has been spent), plus the Budget & Behavior screen in the panel. Plug and unplug are synonyms for license activate and license deactivate.
Shortcuts and integrations
The /up command is installed in Claude Code. Codex uses the installed hooks after you confirm them in /hooks. The panel also exposes /bones, /board and the flow commands on the plans that include them.
Variables and project files
Global state lives in ~/.harnessup/. A project can have its own .harnessup/deploy.json for the deploy screens. Blocks in CLAUDE.md and AGENTS.md are delimited by HarnessUp comments and removed when the matching module is disabled.
Use · 4 min read
Plans and limits
Every plan includes measurement. Each plan includes everything before it, so moving up never turns off what you already use.
Plan
Monthly
Yearly (equivalent/month)
Accounts and projects
Basic FREE
Free
Free
1 Claude Code · 1 project
Pro
US$ 9
US$ 6
1 Claude + 1 Codex · 2 projects
Ultra
US$ 12
US$ 9
2 Claude + 2 Codex + 2 OpenCode · 6 projects
GIGA UNLIMITED
US$ 27
US$ 19
All harnesses · unlimited
Team
—
US$ 1,699/year
10 GIGA licenses
Corporate
On request
—
Negotiated
Sessions are unlimited on every plan. The limit is on active accounts and projects, not usage.
What GIGA adds
KanBan Sync, Code Deck, Flow and Board.
Full Smart Security.
The /bones and /board commands.
The fourth Ez Work block.
Outside the plan, nothing breaks
When an account or project exceeds the plan limit, the event still runs — but every module switches to shadow mode. Nothing is applied, nothing is deleted and the agent proceeds normally. Moved up a plan? Modules start acting again immediately.
7-day trial
The trial gives 7 days of Ultra, is local, asks for no card and applies once per machine. When it ends, the product returns to FREE without losing history or settings.
Billing is handled by Creem as merchant of record, in US dollars and on a recurring subscription. The license floats: one activation per key at a time. Unplug it in the panel, the CLI or /account and plug it into another machine. See plans.
Understand · 5 min read
Privacy and security
Your work content does not leave the machine. What can leave are anonymous counters and hashes — and this page says exactly which.
Daemon guarantees
Listens only on 127.0.0.1, with a 32-byte token per installation.
Bearer token on every route except /health. CORS without wildcards.
The daemon reads and decides, but never executes a command coming from the API.
Never reads, copies, stores or proxies OAuth tokens, API keys or cookies.
What leaves the machine
Data
Leaves?
Details
Code, prompts, command output
Never
Stays on local disk.
File paths
No
Hashed inside the parser and sent as an opaque label.
Bench measurement
Yes, by default
Enums and counters per hour and slice, with a random install_id. Never free text. A test fails if any free string appears.
Panel notifications
Yes
Only {id, event}, with no identifier and no switch.
License validation
Yes
Counters and hashes when there is a license to validate.
Transparency: Bench measurement is on by default on every plan and the switch only exists on paid plans. It sends aggregated numbers, never content. If that does not work for you, the paid plan lets you turn it off.
Vault, quarantine and graph
The vault and quarantine are encrypted at rest, have a 7-day TTL and are never synced. Context Graph is the only place holding repository content and lives in a per-project database at ~/.harnessup/graph/, outside harnessup.db.
No login to measure
You install and measurement starts. There is no login, email or password for FREE. A license is only required for paid modules, and the key is validated locally with cache.
What the product does not do
It does not proxy your Anthropic or OpenAI account.
It does not call the API with your credentials.
It does not send your code to train any model.
It does not retain your transcript content on a server.
The most common stumbles, in the order they appear. If nothing helps, doctor and the logs tell you more than any guess.
Windows warned me when I opened the installer
Expected. The executable has no code-signing certificate yet, so SmartScreen does not know who published it. Check the SHA-256 against SHA256SUMS.txt. If it matches, proceed; if not, delete it and download again. npm and irm do not trigger this warning.
Antivirus flagged the executable
Other than SmartScreen, there should be no alert. If there is, check the hash and the install path. Older versions used a folder layout that confused Defender; the current layout uses one folder per version with a current junction.
`harnessup` is not recognized
Open a new terminal after irm or the installer so PATH is reloaded. With npm, check npm bin -g and make sure the global folder is on PATH. Zip users can call .\harnessup.exe directly.
Hooks installed, but no events arrive
That is exactly the difference harnessup doctor shows. “Registered” is not “called”. Check that the daemon is running with harnessup status, that the port responds at 127.0.0.1:7717, and that the harness was restarted after hook installation.
Port 7717 is taken
The daemon tries ports 7718 through 7727 automatically. runtime.json holds the current port and token. The panel always opens on the right port when you use harnessup open.
Codex does not call the hooks
Codex requires explicit confirmation. Open /hooks in Codex and mark the HarnessUp hooks as trusted. Without that, they stay installed but are never called.
Node is not installed
npm requires Node 18 or newer. Without Node, use irm on Windows or curl on macOS and Linux. Neither path needs an extra runtime.
Mac and Linux
On macOS, use install.sh in Terminal. On Linux, support is outside v1 and the script has been validated on Ubuntu 24.04. There is no .dmg, no notarization and no system service on either platform.
The license will not activate
Check harnessup license status and make sure the key matches the purchase email. The license floats: if it is already active on another machine, unplug it there first. The link in the purchase email also activates automatically.
I deleted `~/.harnessup/` and lost the history
That folder is the local state: settings, license, measurement, vault and quarantine. Deleting it is equivalent to a fresh install. If you need the history, back it up first — there is no server copy.
How to collect a diagnosis
harnessup doctor
harnessup status
harnessup version
Logs live in ~/.harnessup/logs/. To reach real people: help@harnessup.dev.
Reference · Updated Nov 12, 2026
Changelog
What changed in the product, in order. The current version is 1.0.2; installer, irm and zip installs update themselves.
1.0.2Nov 12, 2026
Release 1.0.2: sturdier OpenCode and a fixed Swarm Commander
Improvement
Fix
Security
OpenCode: retries without the variant when the model rejects it, recovers the saved answer when the stream loses it, and adds up the cost of every step.
Swarm Commander: the screen no longer goes blank on the command event, and the council shows the project and a separate count in Subagents Work.
Updates: Scoop, winget, Chocolatey, Homebrew and snap installs are recognized, and the panel shows each manager's update command.
Security: app tools that carry text to another session are handled separately, and the agent's own folders no longer count as writing outside the project.
1.0.1Nov 11, 2026
Release 1.0.1: stable base and a clearer Bench
New
Improvement
Fix
Security
1.0.1 release protocol, with a signed manifest and per-version rollback.
The license's development mode is kept out of the release binary by build tag.
Bench Master: cost per task with every pair, one bar per model, per-model colors, full labels and a non-overlapping axis. Limit raised to 20.
Codex: .zst rollouts without duplicate sessions, plugins adding catalog cost and warm-cache opening.
Website Builder accepts PDF as a reference file, runs OpenCode without a shell and measures tokens instead of cost.
License screen: cards reordered to avoid empty space.
Install: install.ps1 without an internal reference.
PRE-1.0Nov 10, 2026
Deploy, Flow and WhatsApp
New
Improvement
Deploy Accounts gained Resend, Supabase, Sentry and a project name field for one-click deploy.
Flow: BangerMail, Zoho Mail, Mailgun, AgentMail, ShipMail, TrekMail and Winnr in the connector catalog.
WhatsApp as an Actions output and Vercel's Agent Desktop in the shop catalog.
Bench: anonymous token with proof of work and live ingestion enabled.
Ez Work: Subs Director time and turn brakes with confidence.
PRE-1.0Nov 9, 2026
HyperFlow with web search and human approval
New
Improvement
Web search enabled on Codex Base Points.
“Work with OpenCode” toggle on the Base Point, with model and effort.
Human Approval, Local Folder and inbound A2A as configurable connectors.
Hostinger CLI in the Shop and on Base Points.
Bench: cost ignores haiku and zeroed rows; rework vs. model-effort boundary; scoreboard with five.
PRE-1.0Nov 8, 2026
Connectors, CLIs and the Shop
New
Improvement
Eight new connectors with MCP and API, plus AnythingMCP, Agno, Paperclip and Medium in the registry.
Official logos for Shop items that had none.
One-click CLI for Zernio, Clerk, One and Arcade; seven CLIs in HarnessUp Shop.
Connectable IMAP, Notion Calendar through Google and Thunderbird over CalDAV.
Download pages restored for the other Shop items.
PRE-1.0Nov 7, 2026
Content and sending in the panel
New
Improvement
YouTube and Reddit comments in the panel.
Direct sending through ManyChat, Chatfuel and InstantDM.
Instagram integration documentation.
Coming soon: Cursor, OpenCode and Hermes Agent integration; brew, snap, winget and Chocolatey packages; and Microsoft Store approval.
How the product updates itself
The daemon reads https://dl.harnessup.dev/latest.json, verifies the Ed25519 signature and downloads the version's zip. The swap happens on the next daemon shutdown, never mid-session. latest.json is what releases a version to the whole fleet.