MetaHarness CLI
Turns a GitHub repo into a package with its own AI agent: an npx command, an MCP server, memory and a signed release
- 727 ★ on the repository
- Last commit
- MIT
Broad access
Why this level
- Generates executable code: a CLI, an MCP server and config files for Claude Code, Codex, Copilot and other hosts
- The signing and publishing subcommands deal with a signing key, IPFS and your npm account, and the plugin states it validates secrets through GCP Secret Manager
- Network, shell and file-write grants in the generated MCP can be turned on in the policy, after which the agent has them in full
Install
- Works in
Manual install
npx metaharness my-bot --template vertical:coding --host claude-codeCreates a package in the my-bot folder. Swap claude-code for another host, and run npx metaharness --wizard to pick a template.
Access and checks
How we reviewBroad accessThe tool can change data, run commands or handle secrets.
- Generates executable code: a CLI, an MCP server and config files for Claude Code, Codex, Copilot and other hosts
- The signing and publishing subcommands deal with a signing key, IPFS and your npm account, and the plugin states it validates secrets through GCP Secret Manager
- Network, shell and file-write grants in the generated MCP can be turned on in the policy, after which the agent has them in full
This is third-party code. Review the repository files before installing.
What it does
MetaHarness generates an agent harness for a repository that you can publish to npm under your own name. You give it a GitHub repo URL or start from a blank template, and you get a package: a CLI with an npx command, a local MCP server, project-scoped memory, skills built from the actual file layout, an access policy and release checks. It runs in a browser Studio with no install or as npx metaharness. Repository analysis is static: MetaHarness detects Rust, TypeScript and JavaScript, Python and Go from manifests and lockfiles and never runs the project's code. Build and test commands it infers are marked as inferred with execution disabled. `metaharness score` reads a repo and prints a short report card: how well an agent fits, how likely it is to build, how safe the tools are and roughly what a run costs. The generated package connects to ten hosts, including Claude Code, Codex, GitHub Copilot, OpenCode, Hermes and OpenClaw. Its MCP is closed by default: network, shell and file writes are off, dangerous calls need approval, a turn gets at most eight calls, the timeout is 30 seconds and the audit log is on. After that the harness subcommands take over: `harness doctor` and `harness validate` check the package, `harness mcp-scan` looks for risky grants, `harness sbom` produces an SPDX inventory, `harness sign` adds an Ed25519 signature and `harness upgrade` compares with a newer template.
Who it is for. For repository maintainers and engineering teams who want to hand colleagues a ready AI agent for their code behind one npx command and maintain it like any other package.
Good fit when
- An organization has many repositories and each needs its own agent with the same rules, installed with npx @your-org/name
- Before building an agent you want to assess a repo with `metaharness score`: does it fit, how safe are its tools, what does a run cost
- One agent has to work in Claude Code, Codex and Copilot, and each host needs its own config files
- A review needs artifacts: an MCP threat model, an SBOM in SPDX 2.3 and an Ed25519 release signature
- You want to see which agents, skills and MCP tools are recommended for your code, in the browser and without sending data anywhere
Not a fit when
- You need one ready skill or MCP server for a specific task: MetaHarness ships a whole harness, which is too much for a one-off job
- You need a Python agent project skeleton with scenario tests: a separate CLI in the catalog does that
- You must run the agent without Node.js: generation and output rely on npx and npm packages
- You need only verified quality results for agents: the developer calls part of the packages experimental and makes no performance claims for them
Example request
Assess this repository with metaharness score, then create a Claude Code harness for it from the vertical:coding template and check it with harness validateLimitations
- Requires Node.js: every example runs through npx and the result is published as an npm package.
- Part of the packages is experimental. The developer says ARC-AGI-3 and AVO remain private and experimental and that claiming results for them is premature.
- Hosts differ in capability. pi.dev has no MCP by design, Prime Agent gets Python-backed skills without MCP, and with a non-empty deny-list it ships a SANDBOX-REQUIRED.md file instead of enforcing it.
- harness publish --confirm pins the manifest to IPFS, and you publish the package to npm yourself with npm publish.
- The project is large: a generated harness has 21 subcommands, and the npx metaharness --wizard setup is the easiest place to start.
How to disable. MetaHarness itself runs through npx and leaves nothing on the system: it has no background process and no account. The generated package folder deletes like any project folder. If you installed the plugin in Claude Code, remove it through /plugin, and a published package can be withdrawn in npm with the usual tools.
FAQ
How is MetaHarness different from an AI agent framework?
A framework gives a developer parts to build an agent. MetaHarness ships a finished package for one repository, with its own npx command, MCP server, memory, access policy and signed release. The model in such a package is replaceable and the harness stays.
How do I install MetaHarness in Claude Code?
For the plugin run /plugin marketplace add ruvnet/metaharness, then /plugin install metaharness@metaharness, with names taken from the repository's marketplace.json. After install 14 skills are available: creating, checking, scoring and publishing a package. Without the plugin, npx metaharness my-bot --template vertical:coding --host claude-code is enough.
Does MetaHarness run my repository's code during analysis?
No. The analysis and scoring commands only read files. Build and test commands they find go into the report as inferred, with execution disabled.
Do I need a server or an account to use MetaHarness?
No. The browser Studio runs entirely on the client and the CLI runs locally. According to the developer, the project has no account, no hosted backend and no telemetry.
What does harness mcp-scan check?
It is a static scan of the generated package's MCP tools that the developer likens to npm audit for agent tools. It flags shell and network grants, missing audit logs and timeouts, wildcard permissions, unguarded secrets and unpinned dependencies. Any high-severity finding makes the command exit with code 1.
Discussion
No questions or reviews yet.
Related
CLI toolsThe Harness MCP server: one agent starts other Claude Code and Codex sessions on your computer, waits for them, reads their answers and closes them
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 Rust app for macOS, Linux and Windows: skills, MCP servers, plugins, hooks and configs of 17 AI agents in one window, with a security audit
A curated, ranked list of agent harnesses with an MCP server and ready subagents to pick a harness for a task
An open-source agent harness in Rust: GUI, CLI, headless and webapp from one binary, with multiple providers, MCP, skills and agent teams
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
README in short
The repository page is built around the idea of an agent factory: every serious repo deserves its own agent, and MetaHarness mints such packages on request. It opens with a quick start through the Studio, the wizard and npx, then a table of ten hosts that shows what is created for each. After that come MCP with permissions closed by default, 19 ready templates by domain, one-command examples and a table of everyday harness subcommands. The end collects architecture, quality gates, a cost comparison and answers to common questions.