You spent weeks building your Claude Code setup — skills, subagents, project rules, workspace guidance. Then you open Codex, Gemini, or Cursor, and every one of them starts the session half-amnesiac. So you copy-paste folders, rewrite agents in a different format, and watch them drift out of sync forever.

anyagent keeps your Claude Code .claude/ workspace as the single source of truth and projects it into the native formats other agents already know how to read. One canonical setup. Every assistant. No duplicated folders, no copy-paste drift.

npx @saucam/anyagent sync --to codex gemini cursor

That's it. Your Claude skills now show up in Codex's .agents/skills/, your agents become Codex TOML / Gemini Markdown / Cursor .mdc rules, and your CLAUDE.md becomes AGENTS.md, GEMINI.md, and a Cursor workspace rule — all generated from the one workspace you already maintain.

# Tip: install once with `npm i -g @saucam/anyagent` and drop the `npx @saucam/`
# prefix below — the command is just `anyagent`.

# 1. Scaffold a canonical workspace (or use your existing .claude/)
npx @saucam/anyagent init

# 2. See exactly what anyagent can find
npx @saucam/anyagent doctor

# 3. Preview every change without touching disk
npx @saucam/anyagent plan --to codex gemini cursor

# 4. Bridge it
npx @saucam/anyagent sync --to codex gemini cursor

Now the same Claude-origin skill and reviewer agent are discoverable in Codex, Gemini, and Cursor — no rewrite, no second copy to keep in sync.

AI coding tools are converging on the same primitives — skills, agents, rules, commands, hooks, MCP servers — but not on one filesystem layout or manifest format:

The next layer of value isn't another agent. It's the compatibility layer between agents — so your accumulated workflow knowledge belongs to you, not to whichever tool you opened today.

Already using AGENTS.md? anyagent generates it for you — from the CLAUDE.md and .claude/ setup you already maintain — and keeps it in sync, so the standard layout and your canonical source never drift apart.

• Skills are symlinked (relative links, so your workspace stays portable) into each target's native skill folder. Edit once, every agent sees the change instantly. (Cursor has no skill folder, so skills become .mdc rules; Windsurf maps them to native /workflows.)
• Agents are converted into each target's manifest — Codex TOML, Gemini/Hermes Markdown, Cursor .mdc, Windsurf rules — preserving the original instructions.
• Workspace guidance is generated: CLAUDE.md → AGENTS.md / GEMINI.md / .cursor/rules/workspace.mdc / .windsurf/rules/workspace.md / .hermes/WORKSPACE.md.
• Lossy conversions are reported, never silent. When a Claude-specific concept (hooks, tool policies, or Cursor/Windsurf's lack of subagents) has no native equivalent, anyagent preserves it as readable instructions and prints a warn: line. Trust comes from honesty.

$ anyagent sync --to codex
[codex] link skill
  from: .claude/skills/reviewer
  to:   .agents/skills/reviewer
[codex] converted agent
  from: .claude/agents/reviewer.md
  to:   .codex/agents/reviewer.toml
  warn: Claude agent tool and hook semantics are preserved as instructions, not native Codex policy.
[codex] generated guide
  from: CLAUDE.md
  to:   AGENTS.md

3 operation(s), 1 warning(s)

Run it with no install:

npx @saucam/anyagent <command>

Or install globally (the binary is anyagent):

npm install -g @saucam/anyagent
anyagent sync --to codex gemini cursor

Requires Node.js ≥ 20. Zero runtime dependencies.

Shared a .claude/ setup with your team? Make drift impossible to merge. anyagent check exits non-zero the moment a bridged target falls behind the source:

# .github/workflows/agents.yml
name: agents
on: [push, pull_request]
jobs:
  check:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - uses: saucam/anyagent@v1
        with:
          to: codex gemini cursor

✗ out of date — 1 artifact(s) would change:
  [cursor] generated guide: .cursor/rules/workspace.mdc

Run `anyagent sync` to update them.

• One source of truth. Prefer symlinks and generated files over manual copies.
• Native discovery first. Feed each agent the files it already knows how to load.
• No runtime magic. The bridge does its work before the agent starts.
• Lossy conversions must be visible. Never silently drop behavior.
• Local-first. No cloud dependency. No telemetry. Ever.
• Boring filesystem wins. It should feel obvious the moment you see it.

Want a target? Open an issue or read CONTRIBUTING.md — adapters are ~40 lines and share a tested contract.

• Aider and OpenCode adapters
• MCP config normalization across tools
• A hook-compatibility matrix
• Cross-agent command conversion
• anyagent publish — export a portable, shareable bundle

Launching it? There's ready-to-use copy in docs/launch.md. The full vision lives in proposal.md.

Issues and PRs welcome. The codebase is small (~600 lines), strict TypeScript, zero runtime deps, and fully tested. Start with CONTRIBUTING.md.

git clone https://github.com/saucam/anyagent
cd anyagent
npm install
npm run build
npm test

MIT © Yash Datta