Harness MCP
The Harness MCP server: one agent starts other Claude Code and Codex sessions on your computer, waits for them, reads their answers and closes them
- 1,390 ★ on the repository
- Last commit
- MIT
Broad access
Why this level
- Starts other coding agents that run commands and write files in the given folders and in new projects
- May send work to your other machines, where a session acts with that machine's user rights
- Installing Harness means running a script downloaded over the network, and on Linux it installs packages with administrator rights
Install
- Works in
In your terminal, with SkillFoxx CLI
npx skillfoxx add mcp/harness-mcpDetects the agents on your machine, shows the access level and pins the version.
Other ways to install
Access and checks
How we reviewBroad accessThe tool can change data, run commands or handle secrets.
- Starts other coding agents that run commands and write files in the given folders and in new projects
- May send work to your other machines, where a session acts with that machine's user rights
- Installing Harness means running a script downloaded over the network, and on Linux it installs packages with administrator rights
This is third-party code. Review the repository files before installing.
What it does
Harness runs coding agents, Claude Code and Codex, in sessions of their own on this computer or on your other machines. A person sees every session in the Harness app and can join it. The `@autonomous-ai/harness-mcp` package turns this into an MCP server so one agent can manage the others. The server starts with `npx -y @autonomous-ai/harness-mcp` and itself calls `harness mcp` from the installed Harness (the path can be set with the `HARNESS_BIN` variable). The agent gets six tools. `harness_whoami` says who the agent is to Harness and whether it may start sessions now: call it first. `harness_list` shows sessions, their state and who started them, and `harness_start` sends an agent to a task in a given folder or a new project. `harness_wait` waits until sessions stop working, `harness_read` returns a session's last ask and answer, and `harness_close` closes a session. An agent may give work to and close only the sessions it started. If Harness is not installed, the server offers a single tool, `harness_setup`, which explains the installation.
Who it is for. For developers who let a main agent (in Claude Code, for example) hand subtasks to other Claude Code and Codex sessions and want to watch what each of them does.
Good fit when
- A task splits into independent parts and each part should be done by a separate agent in its own folder
- You need to start several sessions and wait for results: `harness_wait` waits until sessions stop working
- You want to watch started sessions and step in from the Harness app instead of reading answers only
- The main agent has to read the outcome of a child session: `harness_read` returns the last ask and answer
- Work runs across several of your machines and sessions must start on different ones
Not a fit when
- Harness is not installed and you cannot install it: without it the server offers only the `harness_setup` tool
- You need agents other than Claude Code and Codex: the server description names only these two
- You use Windows: the Harness installer is described for macOS and Linux
- Subagents inside one Claude Code session are enough: a second layer of session control would only add complexity
Example request
Start two Codex sessions in the api and web folders, give each its part of the task, wait for them to finish and show me their answersMCP
- Transport
- stdio
- Authentication
- not required
Environment variables
HARNESS_BIN- Path to the installed harness command when it is not on PATH
Limitations
The package starts nothing on its own: it wraps the `harness mcp` command and needs an installed Harness. The installer is downloaded and run with `curl -fsSL https://harness.autonomous.ai/cli/install.sh | bash`, and on Linux it installs `tmux` and `curl` through apt-get and asks for a password.
The package version is 0.1.0 and the project is young. Sessions on this computer require no account. Other machines are connected separately and require sign-in. The Harness app runs on macOS or Linux. Node.js 18 or newer is required. The server is distributed under the MIT license and lives in the `cli/packages/harness-mcp` subfolder of the larger OpenHarness repository.
How to disable. In Claude Code run `claude mcp remove harness`, in other clients delete the `harness` block from the MCP settings. Close running sessions with `harness_close` or in the Harness app.
FAQ
How do I start several Claude Code and Codex agents from one agent?
Install Harness, run `harness start` and connect the server with `claude mcp add harness -- npx -y @autonomous-ai/harness-mcp`. The main agent then calls `harness_start` for each subtask and collects results through `harness_wait` and `harness_read`.
What does harness_whoami do and why call it first?
The tool says who Harness takes the agent to be and whether it may start sessions now. That has to be known before `harness_start`, so the author advises starting with it.
Can an agent close someone else's Harness session?
No. An agent hands out work to and closes only the sessions it started itself. It sees sessions begun by you or other agents through `harness_list` but cannot manage them.
What happens if Harness is not installed?
The server still starts but offers one tool, `harness_setup`, which explains how to install Harness. The other five tools appear after installation and a client restart.
Can I install Harness MCP as a plugin?
Claude Code and Codex have a Harness plugin that brings a skill with it. In the repository it is described in `.claude-plugin/marketplace.json` as the `harness` plugin with the MCP server and a skill, and the author points to harness.autonomous.ai/agents for details.
How do I connect Harness MCP to a client without the command line?
Add the block `{ "mcpServers": { "harness": { "command": "npx", "args": ["-y", "@autonomous-ai/harness-mcp"] } } }` to the MCP settings. The server finds the installed Harness itself or uses the path from `HARNESS_BIN`.
Discussion
No questions or reviews yet.
Related
MCP serversTurns a GitHub repo into a package with its own AI agent: an npx command, an MCP server, memory and a signed release
A CLI that keeps an existing agent on a long task: plan a step, run it, verify the result in the real environment and checkpoint progress
A curated, ranked list of agent harnesses with an MCP server and ready subagents to pick a harness for a task
The /better-harness command reviews how an AI agent's work is set up in a project: goals, checks, delivery and memory, and returns a prioritized fix report
A meta-skill for Claude Code that turns a domain description into an agent team and the skills it uses, using six team-architecture patterns
One hundred ready agent teams for Claude Code, each with specialists, an orchestrator skill and domain skills, in English and Korean
Comparison with similar tools
Entries that share the main search query, and how they differ.
| Entry | Official | Authentication | Repository stars | Main difference |
|---|---|---|---|---|
| Harness MCP (this entry) | yes | not required | 1,390 | |
| Claude Squad | no | 8,694 | a terminal app for you: you start several sessions in tmux and worktrees yourself, while Harness MCP lets an agent do it | |
| Orca | no | 109,281 | a desktop environment for parallel agent work in separate git worktrees with a mobile app |
README in short
The main OpenHarness description presents the Harness app for macOS and Linux: parallel agent sessions across machines, a store of ready agents, work with CAD, circuits, robots, games and music, an operating system and an ESP32 device. The MCP package description is short: a table of six tools, the rule about own sessions, Harness installation and client setup. Nearby sits a plugin marketplace manifest for Claude Code and Cursor.