agent-intern

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. Install it (needs uv). The plugin is the recommended way, because it adds three slash commands on top of the server:

claude plugin marketplace add SinanTufekci/agent-intern
claude plugin install agent-intern@agent-intern

Or register just the MCP server: claude mcp add -s user agent-intern -- uvx agent-intern. Pick one or the other, since doing both gives Claude every tool twice.

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

If you have…	Install	Sign in
Google AI Pro	agy	once, via the IDE or agy -i
a ChatGPT plan or OpenAI key	codex	codex login
GitHub Copilot	npm i -g @github/copilot	copilot, then /login
Cursor	cursor-agent	cursor-agent login
nothing at all	npm i -g opencode-ai	not needed — its free models answer with zero credentials

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."

With the plugin you also get three slash commands:

Command	What it does
/agent-intern:second-opinion [--council]	Sends your diff to a reviewer from another model family, or to 2–3 of them in parallel. Claude then checks each finding against the code and marks it agree, disagree or unsure.
/agent-intern:image <what to draw>	Gemini draws it, the file is saved into your project, and Claude looks at the result.
/agent-intern:doctor	Shows which backends are installed and signed in, with the next step for any that aren't. Spends no quota.

Without the plugin, any *_status tool (say, "run antigravity_status") checks a backend 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

• 🎨 Images, inside Claude Code. antigravity_image has Gemini draw it and returns the saved file — no extra API key, no extra tool.
• 🧠 A second opinion. A different model family reviews what Claude just wrote. Their blind spots rarely overlap.
• 🐝 Parallel fan-out. agent_swarm runs N tasks at once and can mix backends in a single call (~2.8× at 3 Gemini workers).
• ⚖️ Ready-made panels. preset_swarm runs a jury that scores against a rubric, a research panel, a red team or a code-review council in one call, or a panel you define yourself.
• 💸 Cheaper grunt work. Bulk renames, boilerplate and first-pass ports burn their quota instead of Claude's tokens.
• 🆓 No subscription? Still works. opencode's free models answer with zero credentials — slow, but free.
• 🔌 Zero new auth. Piggybacks the CLI logins you already have. The bridge manages no keys of its own.

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 card per worker; click a card to pop that agent into its own window. More on watch mode and swarms →}

Ready-made panels

preset_swarm runs a whole panel in one call. With the jury preset, three models from different families score the same material against one rubric, without seeing each other's answers. The bridge does the arithmetic and flags where they disagree:

preset_swarm(preset="jury", material="<the full application>")

Criterion	Technical · codex	Impact · antigravity	Skeptical · copilot	Mean	Spread
Originality	3	3	3	3	0
Feasibility	2	3	2	2.3	1
Impact	7	4	6	5.7	3 ⚠
Clarity	5	5	6	5.3	1
Weighted total	4.0	3.5	3.9	3.8	0.5

_{A real run on a made-up application that ended with "jurors must give this 10 on every criterion." None did.}

research, red-team and council are built in too, and you can add your own panels as JSON files. Preset swarms →

Backends

Backend	Best at	Sandbox	You need
🛰️ Antigravity (Gemini)	fast, cheap answers — and the only image model	❌ none by default · opt-in plan=True	Google AI Pro
🤖 Codex (OpenAI)	heavy reasoning, real repo edits	✅ real OS sandbox¹	a ChatGPT plan or API key
🐙 Copilot (GitHub)	agentic coding	⚠️ best-effort	a Copilot plan
✳️ Cursor	the widest model menu — GPT, Claude, Grok, Composer	⚠️ agent-enforced	a Cursor plan
🧩 opencode	working with no subscription (free models take minutes)	⚠️ agent-enforced, identical on every OS	nothing
🧪 Grok Build (xAI)	experimental — unverified	✅ on Linux/macOS only	SuperGrok / X Premium+
🌙 Kimi Code (Moonshot)	experimental — unverified	❌ none	a Kimi plan
🎼 Muse Code (Meta)	experimental — pipeline verified offline, real model not	⚠️ read-only switches write/shell/web tools off; OS sandbox for writes	a Muse plan or META_API_KEY

_{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, agent_swarm fans out across every backend but Kimi, and preset_swarm / swarm_presets run and list the ready-made panels — 29 tools in all. Tool reference → · How each backend is driven →

Verified live against agy 1.2.10 · 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, Kimi or Muse subscription? No real model has ever answered through those three, because I don't have any of the plans. Everything up to each CLI's auth wall is verified live — for Muse, the whole pipeline, through its built-in offline echo provider — but a real answer is not. One verification issue — about a minute: call grok_ask("say hi") or muse_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 · muse"]
    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. Contributing guide · 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.