Renew Session
ProFinds the turn where continuing costs more than starting over and tells you when the switch pays off. It also tracks how many times the renewal goal was met.
Make it to Friday with room to spare.
Every re-read, giant log and rambling answer weighs on the next round. HarnessUp cuts that waste so your Claude Code and Codex get more done.
26 economy tools. One-click control.
On your machine, with your account and the official agents.
the path of a well-used token
waste filter
your taskrepeated file800-line log“Sure! I’ll…”
HarnessUp at workless context loaded for nothing
your task+ only what helps solve it.
the tokens you save become work :)
Waste has an address. We go get it.
Switch on. Cut.
Watch it happen.
Play with a sample context load. Each switch changes one part of it. In the app, you choose which tools go into action.
Click the switches or press 1 to 4 while the demo is on screen.
12,000tokens in the sample
0tokens avoided
Your task1,200
Large file2,400
Re-read2,400
Long output4,000
Response2,000
> All switches off. Sample context: 12,000 tokens.
The values were chosen to explain the cuts. They are not a savings forecast. In the product, the result depends on the task, the harness and your usage.
This demo only changes the illustration. On Basic FREE, the modules measure in shadow mode; applying the cuts depends on the plan. Check each tool's plan.
ProLong chat.
Long chat.
Growing tab.
The context from the start keeps weighing on the next rounds. Renew Session finds the point where starting over pays off and tells you in how many turns the switch pays for itself.
With /up, you prepare the next session inside the conversation and carry the thread of the work with you.
renewing the session is also moving forward.
cost per roundillustrative drawing
Renew the sessionKeep the same one
Schematic, not to scale. The real point is calculated from your conversation history.
26 ways to
save context.
From the start of the session to the last answer. Each tool cuts an avoidable cost at a different point along the way.
26 tools. The badge shows the minimum plan; each plan includes the previous ones. On Basic FREE, the modules measure without applying cuts.
01
Sessions that know how to restart
/up
ProBrings Renew Session into the conversation. Prepares the switch at the moment it pays off, without losing the thread of the work in progress.
Smart Compress
Basic FREEPrepares the compaction summary from the local history, without calling the model to summarize everything again. On Basic FREE, it measures in shadow mode; applying it follows your plan.
Response Diet
Basic FREEGuides the agent to answer directly, with no preamble or repeating what is already in the code. On Basic FREE, it measures in shadow mode; applying it follows your plan.
02
Reads sized to the task
Code Finder
ProOn the first read of a large file, delivers function names and line ranges. The body comes on request, including in terminal reads. The file stays whole on disk.
Smart Cut
ProDelivers the first page of large files, reports the total size and shows how to ask for the rest. Long logs and data no longer enter the conversation whole, all at once.
Delta Read
ProRemembers which lines the agent has already seen and warns when the file is unchanged. Refuses or shortens re-reads of what is already in context.
Head Limit
ProPuts a cap on unlimited content searches and teaches the agent to ask for the rest. Acts on Grep; file searches by name, like Glob, are left out.
WebFetch Diet
ProAsks for answers of up to 700 words when fetching pages and refuses to fetch the same page again in the session. Requests for full text or for a set length are respected.
03
Smaller outputs. Visible errors.
Terminal Guide
ProSummarizes long test, build and log outputs, keeping the start, the end and the error lines. The full output is stored on the computer.
Output Diet
ProLimits build, test and search output to the start and end of the result. The full content is saved to disk, with a pointer for lookup.
Conduction Write
ProBlocks writes that would leave the file exactly the same, including edits that swap a text for itself. Avoids a call with no change.
No Comment
ProAsks the agent not to comment the obvious and measures how much of the new code still arrives commented. You track whether the guidance is being followed.
First Pass
ProCompiles the project on your machine after an edit and shows the error to the agent before the next turn. The fix goes into the following round.
Way Out
ProWhen a tool fails, points to a fix path that already exists in the project. The hint only appears with evidence and cites the source file.
04
Context that doesn't become baggage
Read Router
ProDelivers the section index of a large context file first. If the agent asks for the file again, it gets the full text. Nothing is deleted.
Hot Context
ProSlices old history into immutable fragments stored in git and searchable. The model reads a short, current view of the context.
Context Graph
ProBuilds a local graph, with no LLM or embeddings. Opens the session with the queue, decisions and measurements, citing file and line. Also flags missing context and duplicate decisions.
Session Protocol
ProWrites into the master file the guidance to open context documents only when the initial summary isn't enough. Avoids rereading everything in each new session.
Memory State
ProThree buttons organize the instructions the agent looks for, teach it to read only the needed section and cap the size of the files it rewrites.
Stature Brake
ProCaps the context file and archives the excess in SafeBox, recoverable with one click. Also hunts for leftovers spread across the project.
Rules Auditor
ProAudits the CLAUDE.md and the rules reloaded in each round. Proposes removing what isn't useful to cut the fixed cost of the whole conversation.
Optimize Harness
ProShows the measured cost of skills, subagents, plugins and servers, with switches to turn each on or off. If a size was not measured, no number is shown.
05
The right work in the right place
Cheap Swarm
UltraDelegates 26 task classes to subagents with the appropriate model and effort. Counts what was delegated, what came back and what the main context stopped paying for.
/cheap
UltraTriggers the Cheap Swarm cast inside the conversation. Each step a subagent can run is delegated; the main conversation decides and gathers the results.
Brings the most recent captures from the Windows screenshots folder into the conversation. You can choose the amount or the screen; without a number, the last three come in.
No tool with that combination.
Try another name or go back to the full catalog.
Try another name or go back to the full catalog.
Cheap Swarm · UltraEven the conductor
Even the conductor
knows how to delegate.
A search and an architecture decision call for different work. Cheap Swarm spreads 26 task classes across models and effort levels, instead of piling everything onto the most expensive model.
You switch it on in the panel. The subagent roster takes the field, and the count shows what was delegated and delivered.
Meet UltraSearch the projectread subagent
Run the teststest subagent
Gather resultsmain conversation
Illustrative flow. The choice depends on the task class.
The proof lives
on your machine.
Your usage has the final word. Start by measuring, see where it weighs and check the effect of the tools on your work.
Understand the measurementFind the spend
Install and run the Deep Scan. It reads the local history and shows where context was spent, conversation by conversation. No login needed to measure.
Choose what makes sense
In the panel, turn on the tools available in your plan. Modules in shadow mode measure without applying changes, including beyond the account or project limit.
Check what happened
Review measurements and events in the panel itself. Compare with your usage and remember: gains vary with task, model, context and frequency of use.
The details
that matter.
Talk to us:
help@harnessup.dev
Does the cut delete parts of my files?
Code Finder and Smart Cut change the section delivered to the agent; the file stays whole on disk. Terminal Guide and Output Diet keep the full output for lookup. Organization tools, like Hot Context and Stature Brake, have their own file process described in the catalog.
What does Basic FREE do in this area?
Modules marked Basic FREE are available for measurement in shadow mode, without applying the cuts. Action depends on the plan you have. Pro unlocks the tools marked Pro; Ultra adds Cheap Swarm and /cheap. Each plan includes the ones before it.
Is the saving the same for everyone?
No. Long conversations, repeated reads and long outputs leave more room for cuts. The effect depends on your work. The demo on this page uses illustrative numbers; the app's measurements come from your own transcripts.
Does it work on the flat-rate plan and the API?
Yes. On the flat-rate plan, avoiding waste helps you make better use of the usage window. On the API, cutting billed tokens can lower the bill. HarnessUp runs the official claude and codex with your account and never reads or stores your token or key.
Do I have to set everything up in the terminal?
No. You turn the tools on and off in the panel, with one click. Commands like /up and /cheap are optional shortcuts inside the conversation, available depending on your plan.
More work.
Breathing room till Friday.
Start by seeing where your tokens go.
Then choose the tools that will cut the waste.
Pro ↗Cut tools
US$ 9/mo7 days free · no card, no signupUltra ↗Pro + Cheap Swarm
US$ 12/mo7 days free · no card, no signup
Monthly prices. Annual: Pro at US$ 6/mo and Ultra at US$ 9/mo, billed annually. See all plans and limits.
Official hooks. Local processing. No proxy.
HarnessUp never reads or stores your token or key.