Getting started

agent-harness-adapter keeps one agent setup in the places coding agents are standardising on, and makes every agent you use read it:

What	Where you write it
Instructions	AGENTS.md
Skills	.agents/skills/<name>/SKILL.md
Hooks	.agents/hooks.json, in Claude Code’s format
MCP servers	.mcp.json
Subagents	.agents/agents/<name>.md, in Claude Code’s format

An agent that already reads a file where it is gets nothing written; Claude Code and Gemini CLI are pointed at it; every other agent gets a copy in its own format. The command calls each agent a harness (--harness claude,codex). How it works explains the details.

Set it up

1. Install the command:

   npm

   npm install -g @six5536/agent-harness-adapter

   cargo

   cargo install agent-harness-adapter

   npm installs a prebuilt binary for Linux and macOS (x64, arm64) and Windows (x64).

2. In your repository, run:

   agent-harness-adapter install

   In a terminal it asks which agents you use (the ones on your PATH are picked) and, when it finds a Claude Code setup, whether to move it into the standard places (see Coming from Claude Code). Then it writes the config, .agents/agent-harness-adapter.toml, and installs. In CI, or to skip the questions, say it up front (without a terminal nothing is migrated unless you add --migrate):

   agent-harness-adapter install --harness claude,codex,cursor

3. Do what install’s notes ask. Some agents load new files only after a step of yours: trust the project and approve the hooks in Codex’s /hooks, trust the folder in Gemini CLI, Copilot, Pi and Devin; restart Codex, OpenCode, Kilo Code and Qwen Code, run /reload in Pi, and start a new Devin session. Harnesses lists them.

4. From now on, edit AGENTS.md and .agents/. Instructions, and hooks on events already in use, apply as soon as you save them. Run agent-harness-adapter install again (from anywhere in the repository) after you add or change a skill, a subagent or an MCP server, or add a hook on a new event.

Keep it in sync

agent-harness-adapter status                      # exit 1 when anything is not current
agent-harness-adapter install --harness gemini    # add an agent
agent-harness-adapter uninstall --harness cursor  # take one out

A file you changed by hand is reported edited and left alone; install --force overwrites it. uninstall takes out what it wrote and never touches AGENTS.md, .agents/ or the config.

In a team, and in CI

• Commit AGENTS.md, .agents/ (with the config and the record, .agents/harness.toml) and the files install wrote for each agent, so everyone gets the same setup without running anything.

• Everyone needs the command. Each agent’s hook entries run agent-harness-adapter. On a machine without it, the hooks cannot start, and like any failing hook they allow: a guard in .agents/hooks.json is then not enforced there. Install it on every machine that runs an agent, or set adapter in the config to a path every machine has.

• Check in CI that nobody forgot to install after a change:

  agent-harness-adapter status   # exit 1 when a copy is out of date or edited by hand

Next

• How it works: sources, copies, the record and the hooks.
• Hooks: one script, every agent.
• Scopes, sharing and edits: your own setup in ~/.agents/.
• For tool authors: ship your tool’s integration the same way.