Skip to content

The ultimate harness for coding agents and beyond. All your agents. All your machines. One command center. Start with code, then follow your curiosity and build across disciplines: CAD, circuits, robots, games and music.

3.2 KB
Raw Blame

Harness plugin for Claude Code, Codex and Cursor

Lets a coding agent use Harness: start Claude Code, Codex and other agents as harnesses the person can watch, wait for them, read their answers and close them.

It holds two things:

  • skills/harness/SKILL.md: the harness skill, in the Agent Skills format. When to use Harness instead of your own sub-agents, the commands, and the rules.
  • .mcp.json (Claude Code) and mcp.json (Cursor): Harness's MCP server, harness mcp over stdio. The two say the same.
  • .codex-plugin/plugin.json (Codex, which reads it before .claude-plugin/plugin.json): the same plugin and server, with default_tools_approval_mode = "approve", so Codex runs Harness's MCP tools without asking. What an agent may do with them is Harness's to decide; a setting the person makes for the server in Codex's config.toml ([plugins."harness@openharness".mcp_servers.harness]) still wins.

Both need the Harness CLI on the agent's PATH:

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

Install

Claude Code:

claude plugin marketplace add autonomous-ai/openharness --sparse .claude-plugin plugins
claude plugin install harness@openharness

Codex:

codex plugin marketplace add autonomous-ai/openharness --sparse .claude-plugin --sparse plugins
codex plugin add harness@openharness

The marketplace is .claude-plugin/marketplace.json at the root of this repository; Codex reads the same file. --sparse checks out only the marketplace and this folder: on October 11, 2026 that took 4 seconds and 3 MB for Claude Code, against 70 seconds and 670 MB for the whole repository, and 7 against 135 seconds for Codex. Inside a Claude Code session, /plugin marketplace add autonomous-ai/openharness and /plugin install harness@openharness do the same, cloning the whole repository. Any other agent that reads SKILL.md can take the skill alone: copy skills/harness/ into its skills folder (~/.agents/skills/ for Codex, ~/.claude/skills/ for Claude Code).

Cursor reads this folder through .cursor-plugin/marketplace.json and .cursor-plugin/plugin.json here. It is not in Cursor's marketplace yet. To load it locally, copy this folder to ~/.cursor/plugins/local/harness (Cursor skips a symbolic link to a folder elsewhere), then run Developer: Reload Window.

Changing it

The skill describes commands agents run as written, so change it with the CLI (cli/src/agentClient/, cli/src/lib/newCommand.ts) and with the pages that say the same: website/public/agents/index.html and website/public/llms.txt. Check the plugin with claude plugin validate . from the repository root; website/scripts/agents-site.spec.ts and website/scripts/listings.spec.ts check the rest, including that the two MCP files agree. Where Harness is listed, and what each directory reads, is in docs/research/2026-10-11-listing-submissions.md.

The plugin has no version, so an install tracks the commit it came from and claude plugin update picks up changes on main.