# Harness — MCP server card for AI agents

Start Claude Code and Codex agents the person can watch, wait for them and read their answers.

Registry name: `ai.autonomous.harness/harness` · category: Memory and agent tools · version 0.4.20 · updated 2026-10-19

## Connect
- npm (stdio): `npx -y @autonomous-ai/harness-mcp`

## Facts
- Repository: https://github.com/autonomous-ai/openharness (1,511 stars)
- Website: https://harness.autonomous.ai/agents

Before you connect a server: read its repository, and give it only the permissions and keys it needs.

## Links
- AgentGid page: https://agentgid.com/mcp/servers/ai-autonomous-harness-harness/


## Harness: start coding agents you can watch

Harness is a free app that runs coding assistants such as Claude Code and Codex, each in its own terminal window you can watch. With this server your assistant can start another assistant on a task, wait for it and read its answer. You see every one in the Harness app and can answer or stop it.

### Where to use it
- A second opinion from another model: While Claude Code builds your feature, it hands a review of the changes to Codex and reads the findings when they are ready.
- Several tasks at the same time: Your assistant splits the work, for example fixing tests in two parts of the project, and starts one helper for each part. It waits for all of them, or for the first one to finish.
- Long jobs that outlast the chat: A slow job, like a long test run, keeps going after your assistant's turn ends, and its answer waits to be read later.
- Use your other computer: Your assistant on the laptop starts a job on your home server or a computer with a strong graphics card, once you have signed in to Harness on both.
- Find what was already decided: Before starting new work, your assistant searches earlier conversations on this computer for what was said about the same problem.

### How to start
- The person installs Harness on a Mac or Linux computer and starts it. No account is needed for harnesses on this computer.

```
curl -fsSL https://harness.autonomous.ai/cli/install.sh | bash
harness start
```

- Add the server to your client.

```
claude mcp add harness -- npx -y @autonomous-ai/harness-mcp
```

- Call harness_whoami first; it says whether the agent may start harnesses and what to ask the person if not.

### License
The app and this server are open source under the MIT license, so you can use them for paid work. Keep the license notice if you pass the code on. The agents you start bill your own Claude or OpenAI plans.

### Good to know
- Every helper it starts uses your own Claude or Codex plan and spends your tokens, so it suits long or parallel work, not quick edits.
- Only you can let your assistant start helpers. You sign in and turn on the setting "Agents can start other agents", which is off by default.
- Once installed, Harness adds its own commands to the permission rules of Claude Code and Codex, so they run without asking you.
- Each helper runs in its agent's auto mode with that agent's own checks on. One assistant may keep at most 4 helpers open at once.
- It works on macOS and Linux only. On Linux the installer adds tmux and curl with apt-get, which asks for your password.

Checked 2026-11-10 against: https://harness.autonomous.ai/agents, https://github.com/autonomous-ai/openharness/blob/f5644bec684e0b98c433acb0253e64674c7ee06e/README.md, https://github.com/autonomous-ai/openharness/blob/f5644bec684e0b98c433acb0253e64674c7ee06e/LICENSE, https://github.com/autonomous-ai/openharness/blob/f5644bec684e0b98c433acb0253e64674c7ee06e/cli/packages/harness-mcp/README.md, https://github.com/autonomous-ai/openharness/blob/f5644bec684e0b98c433acb0253e64674c7ee06e/cli/packages/harness-mcp/index.mjs, https://github.com/autonomous-ai/openharness/blob/f5644bec684e0b98c433acb0253e64674c7ee06e/cli/packages/harness-mcp/package.json, https://github.com/autonomous-ai/openharness/blob/f5644bec684e0b98c433acb0253e64674c7ee06e/cli/packages/harness-mcp/mcpb/manifest.json, https://github.com/autonomous-ai/openharness/blob/f5644bec684e0b98c433acb0253e64674c7ee06e/cli/src/agentClient/mcp.ts, https://registry.npmjs.org/@autonomous-ai%2fharness-mcp

### Mistakes to avoid
- Don't: Start a harness for every small edit → Do: Do quick work yourself (or with your own sub-agents); start a harness only for parallel, long or second-opinion work (Each harness spends the person's tokens.)
- Don't: harness_start with folder set to your own working folder → Do: harness_start with a project name, so Harness makes and trusts ~/harnesses/<name> (A folder Harness did not make makes Claude Code or Codex ask the person to trust it, and the harness waits.)
- Don't: Retry a refused harness_start, or call harness_settings with person_said_yes without asking → Do: Tell the person what was refused; turn a setting on only after their explicit yes in chat (Refusals are final and settings are the person's consent to give.)
- Don't: Treat a harness_wait that returns "Still working" as a failure → Do: Call harness_wait again; one call waits 45 seconds unless timeout_seconds says otherwise (600 at most) (The wait is kept short so clients such as Codex do not end the call; the vendor page still says 120 seconds.)
- Don't: Two harnesses editing the same files → Do: Give each harness its own folder, files or git worktree (Two agents on one file overwrite each other's work.)

### Questions
- Do I need an account? Not for harnesses on this computer. You sign in only to let agents start harnesses, to use your other machines and the phone app.
- What happens if Harness is not installed? The server still starts but offers only one tool, harness_setup, which explains how to install Harness.
- Can my assistant close or take over harnesses I started myself? No. An agent may give work to, and close, only the harnesses it started.
- Is the conversation lost when a harness is closed? No. It is kept and you can resume it.

---
AgentGid's catalog is curated: every entry was checked against the maker's official pages (registry, repository, license and pricing pages) on the date shown; a license value shows the quote it stands on wherever AgentGid has one. Usage numbers come from AgentGid's own scan of public repositories set up for AI coding agents. Claims by makers are reported, not tested by AgentGid.
