AgentMesh

AgentMesh synchronizes project-level AI runtime context across coding tools.

The v0.1 binary is a local-first Rust CLI with bundled Claude Code and Codex adapters. It normalizes project instructions, skills, and subagents into a shared repository model, then renders those entities back into each runtime's native file layout.

Documentation: agentmesh.sh

Install

macOS / Linux:

curl -fsSL https://agentmesh.sh/install.sh | sh

Windows (PowerShell):

irm https://agentmesh.sh/install.ps1 | iex

Upgrade and uninstall options are documented at agentmesh.sh/docs/installation/curl.

For local development, build from source:

cargo build --workspace
./target/debug/agentmesh --help

Quickstart

Prerequisites: a git repository at your project root, and at least one supported runtime present or planned (Claude Code with .claude/ and/or CLAUDE.md, or Codex with .codex/ and/or AGENTS.md).

Preview detection without writing:

agentmesh scan

Initialize AgentMesh from your project root. This detects runtimes, imports entities into the canonical .ai/ model, propagates to other runtimes, installs hooks, and writes agentmesh.lock:

cd /path/to/your/repo
agentmesh init

If AGENTS.md and CLAUDE.md differ, init asks which agent memory file to use as the starting version for initial setup. After setup, sync is bidirectional. For scripts or CI:

agentmesh init --canonical-instructions=AGENTS.md -y

Verify health:

agentmesh status
agentmesh doctor

Commit the shared state teammates need:

git add AGENTS.md .ai/ agentmesh.lock
git commit -m "chore: initialize AgentMesh sync"

Do not commit machine-local hook files (.claude/settings.local.json, .codex/hooks.json). Each teammate runs agentmesh init on their machine. Add .codex/hooks.json to .gitignore.

Codex requires a one-time trust approval before it runs the AgentMesh command hook. After setup, open Codex in the repository and run any tool-backed action; when Codex asks whether to trust the AgentMesh hook command, approve it once. Sync still works via the watcher daemon, Claude hooks, and manual agentmesh sync until then.

Situation	Command
Added a runtime after init	agentmesh install --runtime <name>
Upgraded the binary	agentmesh upgrade
Commit-time drift check	agentmesh install --git-pre-commit
CI pipeline	agentmesh sync --check

Full walkthrough: agentmesh.sh/quickstart

Start, Stop, And Uninstall

To start AgentMesh again for an initialized repository:

agentmesh start -y

This refreshes machine-local AgentMesh state, installs AgentMesh-owned hooks for detected runtimes, and starts the watcher so direct edits to AGENTS.md, CLAUDE.md, and .ai/ files sync immediately. It keeps agentmesh.lock, .ai/, and runtime files such as AGENTS.md intact.

To stop AgentMesh for the current repository while keeping all repository state and AgentMesh installed on this computer:

agentmesh stop -y

This stops the watcher, removes AgentMesh-owned hooks, and clears machine-local cache state. agentmesh.lock, .ai/, and runtime files such as AGENTS.md are left intact so you can re-run agentmesh init later.

To uninstall AgentMesh from the current repository:

agentmesh uninstall -y

This deletes only AgentMesh-owned repository state: agentmesh.lock, .ai/, and agentmesh.config.yaml. AgentMesh never deletes runtime files such as AGENTS.md or CLAUDE.md.

To uninstall AgentMesh from the current repository and this computer:

agentmesh uninstall -y --full

This also removes the agentmesh command from this computer. Runtime files such as AGENTS.md and CLAUDE.md are still retained.

Preview planned stop or uninstall changes without writing:

agentmesh start --dry-run
agentmesh stop --dry-run
agentmesh uninstall --dry-run

More detail: agentmesh.sh/docs/installation/curl.

Development

make ci

The workspace is pinned by rust-toolchain.toml.

Architecture notes for contributors are in ARCHITECTURE.md.