Harness like a boss.

Claude Code at work. Codex at home. Hermes in the cloud. Now all in one command center, yours to run. Start with code, then follow your curiosity into CAD, circuits, robots, games and music.

Download for free Open source

Desktop

Download for Mac Download for Linux

Every computer you own, in one window.

Claude Code, Codex and 12 more agents.

A worktree each, so agents never clash.

⌘N to start, ⌘P to find any agent.

⇧⌘L picks a layout; panes re-tile.

Close the app; your agents keep going.

Harness Store

Claude Codeauth

macbook-pro api main

> refactor auth to use the session cache ⏺ Read(src/auth/middleware.ts) ⏺ Update(src/auth/middleware.ts) ⎿ +18 −31 ⏺ Bash(npm test -- auth) ⎿ 42 passed ⏺ Update(src/auth/session.ts) ⎿ +6 −2 ✻ Cooking… (1m 16s · esc to interrupt)

Pinightly-backup

home-server infra main

› check tonight's backup and summarize the result Reading the backup job logs… Running restic check… ✓ Documents copied to the home NAS · 18m

Codexbilling

office-desktop billing stripe-v2

› migrate billing to stripe v2 • Read billing/charge.ts • Planned 4 steps • Edited billing/charge.ts +88 −52 • Ran npm test billing └ ✓ 31 passed ◦ Working (2m 04s)

OpenCodeeval

gpu-box llm main

┃ run the eval suite on step-4200 + Thought: 117ms → bash python eval.py --ckpt step-4200 mmlu 71.8 +1.2 gsm8k 84.2 +3.0 ▣ Build · deepseek-v4 · 12m 17s

Terminal

$ curl -fsSL https://autonomous.ai/harness.sh | bash$ hn

Control-B, then one key. Like tmux.

Every harness, most urgent first.

Answer agents without opening them.

Any agent, any project, any machine.

Type a task; Harness picks where.

Status and latency of every machine.

hn — studio

Fix flaky login test ⠋studio:webapp ⑂ fix/login-flake #4812

✳ Fix flaky login test

> fix flaky login test

●Reading src/webapp/handler.ts

●Running npm test -- webapp

✓ 41 passed ✗ 1 failed

●The failure is a race in the session refresh — the token is read before the refresh promise settles. Fixing it and re-running.

❯

gpu-box shell ·gpu-box:ml-lab ⑂ main

dev@gpu-box:~/ml-lab$ nvidia-smi --query-gpu=name,utilization.gpu --format=csv

name, utilization.gpu [%]

NVIDIA RTX 4090, 97 %

NVIDIA RTX 4090, 95 %

dev@gpu-box:~/ml-lab$

[studio]0:Fix flaky login test* ⠋1:Add rate limit to t… ?2:Train tokenizer on… ⠋Claude 64%Codex 89%?2✗1⠋2"studio"13:18

Mobile

Download on theApp Store Get it onGoogle Play

Swipe right

Swipe right to see every agent.

Tap the mic

Tap the mic and speak instead of typing.

Tap or say

Tap yes, always or no, or say it.

Swipe left

Swipe left to start a new task.

The ultimate Harness setup.

An always-on display by your keyboard. See which agent needs you, then tap and speak the next task.

Explore the device →

End-to-end encrypted.

Code, keys and keystrokes are sealed on your machine. The relay forwards bytes it can’t read. Harness picks the best path on its own.

Direct.

A WebRTC channel between your machines. No server in the path.

Fast and light.

A native app, not Electron. Close it and your agents keep working. We measure the workflows developers repeat, including their slow tails.

Input to finished frame · 16 terminals open median p95p95 during output

Type → terminal echo11.2 ms74.2 ms

⌘N new harness, input-ready10.6 ms14.4 ms

⌘O open anything12.3 ms16.9 ms

Filter the harness picker10.0 ms14.9 ms

Select a harness → first echo26.1 ms35.9 ms

⌘T new tab13.2 ms18.0 ms

Next tab → first echo27.2 ms42.6 ms

Focus pane → first echo23.9 ms37.3 ms

Zoom a pane55.3 ms326.0 ms

Find query → results10.1 ms14.0 ms

Scroll one viewport40.9 ms139.6 ms

Release build, M2 Max, 120 observations per action. Synthetic harnesses; keyboard, network and display time excluded. Full results

Beyond code: build across disciplines.

Coding agents can build far more than software. Give one a harness and it works with the real tools of a craft while you steer in a live viewer. Every clip below is a real, unedited session.

Design with Blender.

Ask for a lamp and the sliders that matter. Turn them and Blender rebuilds the real geometry. Keep the versions you love.

Circuits with CircuitJS.

Describe a circuit and watch charge move along the wires. Change one resistor, overlay the new scope trace and keep both.

Robotics with MuJoCo.

Build a robot and its controller, then test them in live physics. Shove a paused run with 100 N and watch two futures split.

Games with Godogen.

Your agent builds a game you can play while it takes shape. Miss a jump, rewind, and pin the moment so the agent sees what you mean.

Music with Strudel.

Describe a track and perform it live in Strudel’s REPL. Bring voices in and out, and keep the WAV with its editable source.

Chemistry with RDKit.

Build a molecule from a description, then turn a bond and watch it move along a real energy curve you can reproduce.

Documents with Typst.

Describe a paper or a report and your agent writes a real PDF, updated on every save. Circle a detail and the next draft answers it.

Data with Jev Sheets.

Ask every row a question: reviews, surveys or tickets come back with answers and probabilities. Test a wording on a few rows first.

Domain-Specific Harnesses.

Each one turns a coding agent into a specialist. You chat on one side; the part, the board or the game takes shape on the other.

See all 49 harnesses →

Claude CodeWork with Claude directly in your codebase

CodexCoding agent that runs in your terminal

CursorShip code with agents, right from your terminal

OpenCodeOpen source AI coding agent

PiTerminal-based coding agent

HermesOpen-source AI agent that grows with you

Command CodeCoding agent built for open models

DevinAI software engineer

Muse CodeCoding agent for complex coding workstreams

AmpCoding agent and dev environment

Kilo CodeOpen source AI coding agent in IDE, CLI and cloud

Grok BuildCoding agent that runs right from your terminal

AntigravityNext-gen agent platform

GitHub CopilotRun a GitHub-native agent in your terminal

Harness Monitorhtop for your harnesses: pause the idle ones, resume any of them

Machine MonitorEvery computer you own, on one live map

Before you start.

How do remote harnesses stay responsive?

Harness uses a direct WebRTC connection when your network allows it, so terminal traffic travels between your machines without a relay in the data path. It sends terminal input and output, rather than streaming a remote desktop. The actual response time still depends on your network and host machine. Your agent’s model speed is separate.

What if a direct connection is blocked?

Harness can carry the encrypted connection through Cloudflare TURN or use a WebSocket relay while WebRTC connects. It can switch to a direct connection when one becomes available. You don’t need to open inbound ports or configure a VPN.

Can a relay read my terminal traffic?

Remote terminal traffic is encrypted end to end, with the keys held by your connected machines. Relays forward the encrypted data. That protects the terminal stream; it doesn’t mean the service has no account or connection metadata.

What keeps a harness running?

The agent process runs on its host machine independently of the Harness window. Closing the app or losing your connection doesn’t stop work on an awake remote host. Reconnect to the same harness when you’re ready.

What if the machine running an agent sleeps or restarts?

A harness still needs a running host. A sleeping machine can’t keep doing work. After a reboot, the daemon restores the panes and resumes their harnesses once Harness starts again. For jobs you want to leave running, use an always-on home server, office machine, or cloud server.

Where do my agents and files actually live?

Each agent runs on the machine hosting its harness and works with the files and tools available there. Putting harnesses side by side in Harness doesn’t copy their projects between machines. Your agents continue to use their own model providers and credentials.

How do I connect a remote machine?

Install Harness on it (on a server without a screen, the CLI), sign in with the same Harness account and set a remote password. Then choose Machines → Link Machine in the app and enter that password. The link is end-to-end encrypted, with no SSH keys and no open ports.

Can I use Harness with just one computer?

Yes. Start with local agents on your Mac or Linux computer. Remote machines are optional.

Do I keep my existing agent accounts?

Yes. The agents run with your own accounts or API credentials. Harness gives you a place to work with them across your machines.

What does the free download include?

The Harness desktop app is free to download. Bring your own agent accounts; agent subscriptions, model usage, and any rented servers are separate.

Do I need the Harness device?

No. The desktop app works on its own. The device is an optional companion for controlling your harnesses from your desk.

Every agent, every machine

fix-loginstudio:web⑂ fix/login-refreshrefactor-db asking

> sessions log out after an hour, find out why

I'll reproduce it with the auth tests first.

Bash(npm test -- auth)

⎿41 passed, 1 failed

✗ refresh › renews before the token expires

Read(src/auth/refresh.ts)

⎿Read 88 lines

The refresh timer reads expires_in as minutes, but it's seconds, so the token runs out long before the refresh.

Update(src/auth/refresh.ts)

⎿Updated src/auth/refresh.ts with 1 addition and 1 removal

Bash(npm test -- auth)

⎿42 passed

Fixed. Want a test that covers an expired token?

App Store Google Play

Harness like a boss.

Desktop

01Every machine, side by side.

02Every agent, side by side.

03Each agent on its own branch.

04Keyboard first.

05Your layout, your way.

06Agents keep working without you.

Terminal

01One prefix for everything.

02Find any harness.

03Answer the agents that are waiting.

04Start a new harness.

05Send a task without choosing where.

06See all your machines.

Mobile

01Find any agent

02Talk instead of typing

03Reply when an agent asks

04Start a new task

The ultimate Harness setup.

End-to-end encrypted.

Fast and light.