agent-intern

Give Claude Code an intern.

Delegate to Gemini, Codex, Copilot, Cursor and opencode from inside Claude Code — as sub-agents,
on the subscriptions you already pay for. Text answers, image generation, real repo work, parallel swarms.

Quick start · What it's for · Backends · Security · Docs

Claude Code is one model on one quota. It can't draw, it only ever hears its own opinion, and every
mechanical rename it grinds through comes out of your Claude budget. Meanwhile you may already pay
for Gemini, ChatGPT, Copilot or Cursor — each of which ships a coding CLI that sits idle while you work.

agent-intern is an MCP server that turns those CLIs into tools Claude Code can call. Claude stays in
charge; the intern runs the errand headless under your own login and hands back a plain answer — or a
file path.

Quick start

1. Register the server (needs uv):

claude mcp add -s user agent-intern -- uvx agent-intern

2. Install at least one backend CLI and sign in once. Any one of them works on its own:

3. Restart Claude Code and just ask. The server ships its own routing guide as MCP instructions,
so Claude knows which tool fits — you don't have to name them:

"Ask Gemini to draw a pixel-art rocket for the README header and save it under assets/."
"Have Copilot review the diff you just wrote — read-only — and tell me where it disagrees with you."
"Summarise each of the six files in src/handlers in parallel with a swarm."

Any *_status tool (say, "run antigravity_status") checks a backend's setup without spending quota.

[!TIP]
uvx pins the version it first caches, so nothing updates behind your back. Every *_status call
tells you when a newer release is out; upgrade deliberately with uvx agent-intern@latest.
Other install paths, from source included →

What it's for

Watch it work

Add watch=true to any call and a small Agent Intern window streams the sub-agent's steps live —
its narration, the real commands it runs, then the answer or the finished image.

a text ask	antigravity_image — image inline

_{agent_swarm(..., watch=true) — one row per worker; click a row to pop that agent into its own window. More on watch mode and swarms →}

Backends

_{¹ On Windows, codex 0.149.1's sandbox refuses every command — reads included — so a sandboxed Codex
can't see your files there. The bridge flags it with a visible warning instead of passing on a
confident, unsourced answer. Details →}

Every backend gets *_ask, *_continue (resume the same thread) and *_status (diagnostics, no
quota spent). Antigravity adds antigravity_image and antigravity_image_swarm, and agent_swarm
fans out across every backend but Kimi — 24 tools in all.
Tool reference → ·
How each backend is driven →

Verified live against agy 1.1.25 · codex-cli 0.149.1 · copilot 1.0.80 · cursor-agent 2026.07.23 ·
opencode 1.18.29. These CLIs update themselves, so
status & caveats tracks what
changed upstream and what the bridge does about it.

[!IMPORTANT]
Have a Grok or Kimi subscription? Neither backend has ever completed an authenticated round-trip,
because I don't have either plan. Everything up to each CLI's auth wall is verified live; behind it,
nothing is. One verification issue
— about a minute: call grok_ask("say hi") — is the most useful contribution you can make.
Details →

How it works

flowchart LR
    U([You]) --> CC([Claude Code])
    CC -- "MCP tool call" --> B["agent-intern<br/>(MCP server)"]
    B -- "headless, one-shot,<br/>your own login" --> CLI["agy · codex · copilot · cursor-agent<br/>opencode · grok · kimi"]
    CLI -- "answer or file" --> B
    B -- "plain text" --> CC

Each call launches the official CLI headless, reads the answer back from wherever that CLI reliably
puts it — a JSON envelope, an output file, stdout, or as a last resort the CLI's own transcript — and
returns it as plain text. *_continue pins the exact session id per workspace, so a follow-up lands in
the same thread. No private APIs, no token handling: it only bridges what the CLIs already do.

Security

[!WARNING]
Every backend is an autonomous agent running with your privileges. Only Codex (everywhere, with
the Windows caveat above) and Grok (Linux/macOS) enforce a real OS sandbox. Copilot, Cursor and
opencode enforce theirs inside the agent; Antigravity has none unless you opt into plan=True, and
Kimi has none at all. workspace is a starting directory, not a boundary. Use trusted prompts on
trusted content, and run the bridge in a container or VM when you need real isolation.
What each sandbox actually enforces →

Docs

• Setup & requirements — install from source, per-backend prerequisites, *_BIN overrides
• Tool reference — every tool, argument and default
• Backends in depth — answer paths, session resume, models & auth, the experimental backends
• Watch mode & swarm
• Security
• FAQ — terms of service, cost, which backend when, making Claude offer to delegate
• Status & caveats — version-by-version compatibility notes
• Changelog

Contributing

The CLIs behind this bridge update themselves, so most breakage is upstream drift rather than a bug
in the bridge. A report that includes the relevant *_status output usually pins it down in one go.
Open an issue ·
Start a discussion · Developed on Windows —
confirmations from macOS and Linux are very welcome.

Community & acknowledgments

Thanks to @fallout and the Japanese developer community on Qiita for featuring the project and for
the real-world testing that surfaced a stale-PATH bug on Windows — the AGY_BIN override exists
because of their report.
Hybrid setup guide (Claude Code × Antigravity CLI) ·
Quick installation guide

License

MIT. Do whatever you want with it.