agents-anywhere

Your AI agent configs, skills, and instructions — on every device.
One config repo. Every agent. Every machine.

---

You've spent hours perfecting your CLAUDE.md, building custom skills, tuning your settings. Then you open your laptop and none of it is there. Or you switch from Claude Code to Codex and start from scratch.

agents-anywhere keeps your agent setup in a git repo and symlinks it to every agent on every machine.

Quick Start

# New setup — one command does everything
npx agents-anywhere init

# New device — clone your existing config
npx agents-anywhere init --from https://github.com/you/agents-anywhere-config.git

That's it. init detects your installed agents, picks a primary, copies configs, imports MCP servers, creates symlinks, and optionally pushes to a private GitHub repo.

Supported Agents

Agent	Instructions	Skills	Other portable files
Claude Code	CLAUDE.md	skills/	settings.json, keybindings.json, commands/
Codex CLI	AGENTS.md	skills/	config.toml
OpenCode	AGENTS.md	skills/	opencode.json
Gemini CLI	GEMINI.md	skills/	settings.json, commands/
Cursor	rules/	skills/	—
Windsurf	rules/	skills/	memories/
GitHub Copilot CLI	copilot-instructions.md	skills/	—
Amazon Q Developer	rules/	skills/	—
Kiro	steering/	skills/	—
Antigravity	GEMINI.md	skills/	—

Cursor, Windsurf, and Amazon Q only support project-level rules — global instructions are not synced for these agents.

Planned: Cline, Roo Code, Kilo Code, Amp, Augment, Zed, Trae, Continue.dev.

Device Sync

# Machine A — you changed your CLAUDE.md and added a skill
agents-anywhere push

# Machine B — get the changes
agents-anywhere pull
# post-merge hook auto-runs: link + mcp sync

No cloud, no accounts — just git.

What Gets Synced

agents-anywhere-config/           # your git repo
├── claude-code/
│   ├── CLAUDE.md                 # → ~/.claude/CLAUDE.md (source of truth)
│   ├── settings.json             # → ~/.claude/settings.json
│   ├── skills/                   # → ~/.claude/skills/
│   └── commands/                 # → ~/.claude/commands/
├── codex/
│   ├── AGENTS.md                 # → symlink to claude-code/CLAUDE.md
│   ├── config.toml               # → ~/.codex/config.toml
│   └── skills/                   # → ~/.codex/skills/
├── opencode/
│   ├── AGENTS.md                 # → symlink to claude-code/CLAUDE.md
│   └── skills/                   # → ~/.config/opencode/skills/
├── mcp.json                      # normalized MCP config
└── agents-anywhere.json          # manifest

Your primary agent's instructions (e.g., CLAUDE.md) become the source of truth. Other agents get symlinks with the correct filename. Edit once, synced everywhere.

MCP Normalization

Every agent has a different MCP config format. Write one mcp.json, get the native format for each agent.

init auto-imports your existing MCP servers from all installed agents and deduplicates them.

{
  "servers": {
    "github": {
      "transport": "stdio",
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-github"],
      "env": { "GITHUB_TOKEN": { "$env": "GITHUB_TOKEN" } }
    }
  }
}

$ agents-anywhere mcp sync

✓ Claude Code     — wrote ~/.claude/.mcp.json
✓ Codex CLI       — merged into ~/.codex/config.toml
✓ Cursor          — wrote ~/.cursor/mcp.json
✓ Gemini CLI      — merged into ~/.gemini/settings.json
...

Each agent gets its native format — correct root keys, env var syntax, transport naming.

MCP Format Reference

Agent	Output	Root key	Env syntax
Claude Code	~/.claude/.mcp.json	mcpServers	${VAR}
Codex CLI	~/.codex/config.toml (merged)	[mcp_servers.*]	env_vars array
OpenCode	~/.config/opencode/opencode.json (merged)	mcp	{env:VAR}
Gemini CLI	~/.gemini/settings.json (merged)	mcpServers	${VAR}
Cursor	~/.cursor/mcp.json	mcpServers	${env:VAR}
Windsurf	~/.codeium/windsurf/mcp_config.json	mcpServers	${env:VAR}
GitHub Copilot CLI	~/.copilot/mcp-config.json	mcpServers	${VAR}
Amazon Q Developer	~/.aws/amazonq/mcp.json	mcpServers	${VAR}
Kiro	~/.kiro/settings/mcp.json	mcpServers	${VAR}
Antigravity	~/.gemini/antigravity/mcp_config.json	mcpServers	${VAR}

Commands

Command	Description
init [dir]	Detect agents, copy configs, import MCP, link — full setup
init --from <url>	Clone an existing config repo and link
push	Stage, commit, and push config changes to remote
pull	Pull config changes (post-merge hook re-links)
link [agent]	Symlink configs to agent directories
unlink [agent]	Remove symlinks, restore backups
status	Show link status per agent and file
agents	List all known agents with install status
enable <agent>	Enable an agent in the manifest
disable <agent>	Disable an agent in the manifest
mcp sync	Generate per-agent MCP configs from mcp.json
mcp add <name>	Add an MCP server interactively or with flags
mcp remove <name>	Remove an MCP server
mcp list	Show all configured MCP servers
mcp diff	Preview what mcp sync would change
doctor	Diagnose broken symlinks, credentials in repo, stale configs
validate	Validate bundled agent definition schemas
export	Generate a standalone install script

link, unlink, and mcp sync support --dry-run.

How It Works

Each agent is defined by a JSON file — no TypeScript code needed:

# Add support for a new agent = add one JSON file
agents/my-agent.json

The definition declares where configs live, what files are portable, what to ignore, how MCP is formatted, and whether the agent supports global instructions.

See DEVELOPMENT.md for the full schema and contribution guide.

License

MIT