Force Test
Code changed and no test passed after the change? The session doesn't end, and the agent goes back to fix it.
3 blocks no green test after 3 blocks: the session closesYou call the shots.
The agent obeys.
A spend cap, working hours, rules that don't evaporate halfway through the conversation and gates the session has to pass before it closes. Control goes back into your hands.
25 control tools, from Basic FREE to GIGA UNLIMITED.
Each one switches on and off with one click in the panel.
25 control tools17 ready-made conventions in ToolBox4 work intervals per day5 git locks, each with its own switch
Full control
in your hands!
This is Stop Loss in pocket size. Pick the unit, set the limit and release the agent. When the session passes the cap, the next tool call is refused and everything closes in order, without you watching.
Session limit
Choose the unit the brake counts in.
spendUS$ 0.00
tokens0k
turns0
- no calls yet. Hit Release the agent.
Simulation with illustrative costs that grow every turn, like a real conversation. Shortcuts with the panel in focus: G releases the agent, R restarts, 1 2 3 switch the unit.
Budget Cap guides. Stop Loss brakes.
Budget Cap sets spend and usage guidance per task, session or project, and tells the model how to work within it. It's an agreement with the model, with no automatic cut. For the actual cut, Stop Loss is the one that acts.
ToolBox · Rule Checker · Pro
The rule that sticks.
The agent forgets the agreement on turn 12. HarnessUp delivers the rule at the right moment and checks that it was followed.
ToolBoxcost per turn: 0 tokens
Sample of conventions. The panel has 17, each with its per-turn cost measured upfront. Here the math is approximate: characters ÷ 4.
When each rule arrives
Texts delivered to the model: 0Rule Checker alerts: 2
- Init ConventionPro
One text or up to 6 files delivered once, when each new session opens, globally or per project. A resumed, compacted or forked session doesn't get them again.
- Stick PromptUltra
The rule the agent keeps forgetting comes back with every prompt you send, until you remove it.
- Anti-fadingUltra
A silent reminder every N turns, from 1 to 100, up to 1,000 characters. Applies to one session, one project or all of them, and never shows up in the conversation.
- Rule CheckerPro
At the end of each response, checks the ToolBox rules that leave a trace: a file cited without being opened, a loosened test, a type escape hatch. Found a failure, it tells you. Blocks nothing.
MondayWorking Hours · 3 of 4 intervals
10:00
Within working hours. Tools are accepted.
In the demo, each minute of break passes in 1 second.
Working Hours · Scheduled Break · Pro
The agent clocks in, too.
Set when the agents may work. Outside those hours, every tool is refused, in any session open on the computer.
Each day accepts up to 4 intervals: morning, afternoon, a lunch breather and an evening window. Scheduled Break plans pauses with an on-screen countdown. During a break, tools are refused and new requests wait until the time runs out.
See the Pro planNo test, no phase close.
Four gates at the end of the road. The session only closes after it passes them, and the agent goes back to fix what's missing. With an attempt limit, so the brake never turns into a prison.
Enforce TDD
Test before code. Editing a code file is only released after a test has been written or run in red since the last request.
docs and config pass straight through; the TDD rule applies to every requestForce Lint
Changed code only ships with the project's linter passing. The agent runs it, fixes what it flags and is banned from loosening the config.
no turning off lint rules to close the taskDoc Now
Edited code? Updates the project docs, such as TODO.md and DECISIONS.md, before the turn ends.
3 blocks gives up after 3 blocks in the same turnEach gate is a separate switch in the panel.See the Ultra plan
Main is yours. The agent asks permission.
Locks the most sensitive git actions in every session. Each group has its own switch: release one without opening the others.
Click a command.
See what Git Admin does with it, based on the locks beside it.
Waiting for the agent.
Pick a command above.
Subs Director · Ultra
A subagent with a time to come back.
Every subagent gets a budget of turns and time. Nobody blows the limit or comes back empty-handed.
subagent: test search · 20 turnsillustrative example
- 50%half the budget: warning sent
- 75%started reading outside the task: course corrected
- 90%time to wrap up: report requested
Swarm Commander
UltraWatches the session in real time. When the main model would do alone work that could be delegated, it hands the task to the Cheap Swarm crew. Only acts with Cheap Swarm on.
No Subagents
ProThe opposite, when you need it: no subagent opens. The call is refused and the agent does the work in the conversation itself, with no unexpected delegation on the bill.
/watch
UltraPuts an agent on the other open sessions, without interfering with their work. At the end, it delivers a verdict on what it saw.
Anti-Collision
GIGA UNLIMITEDTwo sessions on the same file or the same folder? The bigger model wins. The smaller one waits, parked, spending no tokens, until its turn is released.
Anti-Collision
A queue at the file's door.
With several sessions in the same project, nobody overwrites anyone's work. The waiting one stays truly parked, running no turns and burning no window.
Cartridges for every agent quirk.
Small tweaks that cut wasted rounds, pick the right effort and put the right material in front of the model at the right moment.
Reasoning Advisor
Suggests the model and the effort the task calls for, no guessing. The model comes in the first prompt, when switching doesn't yet break the cache. After that, only the effort.
ProeffortProductivity Mode
Holds back one-line requests and reminds you to bundle 3 or 4 into one message. Resent within 30 minutes? Goes straight through.
Ultra1 time per textDo Not Reply
Hides the Claude Code button that dismisses the next-step suggestion. That click would make the model reread the whole context just to hear "no need".
Pro0 wasted roundsAttach Skill
Attaches up to 8 of your skills to a moment in the session and puts them in front of the model when it arrives. The catalog gathers skills from you, from plugins and from open projects, with the description in view.
Proup to 8 per ruleAttach Files
Attaches key files from a project to a moment in the session and points the model to them at the right time. Great for a short task inside a big project, without explaining again where everything is.
Profile:momentToolBox
17 ready-made conventions for the agent to work with discipline, switched on one by one, each with its per-turn cost in view.
Prosee above ↑HarnessUp OFF
Turns off everything HarnessUp does for a while, so you can work with the bare agent without uninstalling anything. When the chosen time is up, it comes back on its own. Good for comparing with and without, or for debugging a problem.
everything onThe right control for your plan.
Each plan includes the tools of the ones before it. Start with a spend cap and tighten the rules as the project grows.
Compare pricing and plansBasic FREE
HarnessUp OFF, to compare the agent with and without the layer.
↗Pro
ToolBox, Budget Cap, Stop Loss, Do Not Reply, Rule Checker, Init Convention, Reasoning Advisor, Attach Skill, Attach Files, Working Hours, Scheduled Break, Git Admin, No Subagents and Force Test.
↗Ultra
Adds /watch, Swarm Commander, Productivity Mode, Stick Prompt, Anti-fading, Subs Director, Enforce TDD, Force Lint and Doc Now.
↗GIGA UNLIMITED
Everything in Ultra, plus Anti-Collision for several sessions in the same project, with unlimited accounts and projects.
↗Local.on your machine
Official hooks, your account untouched.
HarnessUp runs the official claude and codex with your account and never reads or stores your token or authentication key. The locks and reminders use official hooks from Anthropic and OpenAI, with no proxy or API interception, and a p95 latency budget under 50 ms.
Before you hit the brake.
What each tool holds, refuses or reminds, with no extra promises.
help@harnessup.devDoes Stop Loss cut the session in the middle of an edit?
No. When the session passes the limit in US$, tokens or turns, it refuses the next tool call and the session closes in order. Whatever was in progress finishes first.
What's the difference between Budget Cap and Stop Loss?
Budget Cap gives the model spend guidance per task, session or project, and sets how it works within them. It's guidance, with no automatic cut. Stop Loss is the brake that actually stops the session.
Don't Stick Prompt, Anti-fading and Init Convention do the same thing?
They differ in timing. Init Convention delivers once, when a new session opens. Stick Prompt repeats with every prompt you send. Anti-fading sends a silent reminder every N turns, from 1 to 100. You can combine all three.
Can Force Test lock the session forever?
No. After 3 blocks without a green test, the session closes. Doc Now also gives up after 3 blocks in the same turn, and Productivity Mode lets through the same request resent within 30 minutes.
Does Working Hours apply to every session?
Yes. Outside the set hours, every tool is refused in any session open on the computer. Each day accepts up to 4 intervals.
Are Control and Protection the same front?
Control defines how the agent works: cap, schedule, rules and gates. Protection handles undoing damage and closing doors such as network and MCP. See Protection ↗
Take the controls. Let the agent play.
Turn on cap, schedule and rules in your plan. The metering layer is free and needs no login, so you can see where the agent spends before you decide.