The Problem

AI coding CLIs each store configuration differently. Claude Code uses .claude/commands/ and .mcp.json. OpenCode reads .agents/ directly and uses opencode.json. Codex reads .agents/ directly and writes MCP servers to .codex/config.toml. Optional adapters such as Cursor, Copilot, RooCode, and Cline have their own file formats too.

When you use multiple tools — or your team does — you end up maintaining the same rules, commands, and MCP server configs in 3-5 different places. Add a new MCP server? Edit 4 config files. Update a coding standard? Copy-paste across directories. Onboard a new tool? Manually recreate everything.

In monorepos it's worse. Each service needs its own tool configs, but shares org-wide standards. Without a sync layer, config drift is inevitable — your frontend team's Cursor rules diverge from the backend's Claude rules within weeks.

The Solution

AgentSync is a CLI that reads one config (.agents/agentsync.toml) and writes to every tool's native format. Skills, commands, and MCP servers defined once, synced everywhere. Non-interactive, deterministic, CI/CD friendly.

┌─────────────────────┐
│  .agents/           │
│  ├── agentsync.toml │──── agentsync sync ────┬──► .claude/skills/*.md
│  ├── skills/*.md    │                        ├──► .claude/commands/*.md
│  ├── commands/*.md  │                        ├──► .mcp.json (Claude)
│  └── agents/*.md    │                        ├──► opencode.json
│                     │                        ├──► .codex/config.toml
│  Presets:           │                        └──► more CLI adapters
│  github:company/std │
└─────────────────────┘

Tools that read .agents/ natively (OpenCode, Codex, Gemini, Amp, Goose, and others) need zero file copies -- AgentSync just ensures the source directory is correct. Tools that need native files get generated outputs in their tool directories.

Install

npm install -g agentsync

Quick Start

# Initialize in your project (non-interactive by default)
agentsync init

# Sync to all enabled tools
agentsync sync

# Preview changes without writing
agentsync sync --dry-run

# Diagnostics
agentsync doctor

# Remove all synced files
agentsync clean

This creates .agents/agentsync.toml:

tools = ["claude", "opencode", "codex"]

extends = ["github:company/standards"]

[mcp.github]
command = "npx"
args = ["-y", "@modelcontextprotocol/server-github"]

# Add an MCP server to config
agentsync config add mcp github --mcp-config '{"command":"npx","args":["-y","@modelcontextprotocol/server-github"]}'

# Remove an MCP server
agentsync config rm mcp github

# List configured MCPs
agentsync config ls mcp

# Dump resolved config
agentsync config show

MCP servers defined in [mcp.*] are enabled by default -- defining a server is enabling it. No separate mcp_enabled list needed.

Pull skills, commands, and MCP configs from GitHub repos or local directories:

extends = [
  "github:company/standards",
  "fs:./local-presets",
]

Presets are namespaced to prevent conflicts (company--typescript.md in tool outputs). GitHub presets are fetched fresh during sync; there is no persistent preset cache or --pull flag in the v1 beta CLI.

AgentSync walks from CWD to git root, merging every .agents/agentsync.toml it finds:

my-monorepo/
├── .agents/agentsync.toml          # Org-wide: tools, shared MCPs
├── frontend/.agents/agentsync.toml # Frontend overrides
└── backend/.agents/agentsync.toml  # Backend overrides

Running from frontend/ merges both configs. Tools lists replace (deepest wins), MCP servers merge per-key (deepest wins).

Define multiple configurations in one file, selected at sync time:

[profiles.frontend]
tools = ["claude", "opencode"]
mcp = ["storybook", "figma"]

[profiles.backend]
tools = ["claude", "codex"]
mcp = ["postgres"]

[profiles.ci]
tools = ["codex", "claude"]
env = "CI"

Profile fields like mcp, skills, and extends use filter semantics -- they restrict the base config to the listed items, rather than accumulating.

agentsync sync --profile frontend
# or
AGENTSYNC_PROFILE=ci agentsync sync

Tool Tiers

Validated beta support means the adapter is part of the maintainer release-validation set.

| Validated CLI | Skills | Commands | MCP | Method | |---------------|--------|----------|-----|--------| | Claude Code | yes | yes | yes | File copies (.claude/) | | OpenCode | yes | yes | yes | Reads .agents/ directly | | Codex | yes | no | yes | Reads .agents/ directly | | Gemini | yes | no | yes | Reads .agents/ directly | | Amp | yes | yes | yes | Reads .agents/ directly | | Goose | yes | no | yes | Reads .agents/ directly | | Aider | yes | no | no | AGENTS.md only | | Amazon Q | yes | no | yes | Reads .agents/ directly | | Augment | yes | yes | yes | Reads .agents/ directly | | Kiro | yes | no | yes | Reads .agents/ directly | | OpenHands | yes | no | yes | Reads .agents/ directly | | Junie | yes | no | yes | Reads .agents/ directly | | Crush | no | no | yes | MCP only | | Kilocode | yes | no | yes | Reads .agents/ directly | | Qwen | yes | no | yes | Reads .agents/ directly |

Optional adapters are still supported by the config schema and sync engine, but they are not part of the default validated beta target set.

| Optional Adapter | Skills | Commands | MCP | Method | |------------------|--------|----------|-----|--------| | Cursor | yes | yes | yes | File copies (.cursor/) | | Copilot | yes | no | yes | File copies (.github/) | | RooCode | yes | yes | yes | File copies (.roo/) | | Cline | yes | no | no | File copies (.clinerules/) |

git clone https://github.com/baranovxyz/agentsync.git
cd agentsync
pnpm install
pnpm build
pnpm test

See CONTRIBUTING.md for contribution guidelines.

Links

• npm package
• GitHub repository
• Configuration guide
• CLI reference
• Architecture
• Changelog

License

MIT