polyclaude 0.3.0

Claude Code is a great terminal coding agent — but it only talks to Anthropic's models. polyclaude lets you keep the exact Claude Code experience (its UI, tools, and agent loop) while the actual model behind it is GPT, Gemini, a Groq model, an OpenRouter model, or a local Ollama model.

It works by running a tiny local proxy that speaks Claude Code's API on one side and your provider's API on the other, translating between them on the fly. Claude Code itself is never modified.

Claude Code  ⇄  Anthropic Messages API  ⇄  [ polyclaude ]  ⇄  OpenAI / Gemini / …

Install

pipx install polyclaude      # recommended
# or: uv tool install polyclaude
# or: pip install polyclaude

You'll also need Claude Code installed (claude on your PATH) and an API key for whichever provider you want to use.

Use

export GEMINI_API_KEY=...        # or OPENAI_API_KEY, GROQ_API_KEY, …
polyclaude --gemini              # Claude Code, powered by Gemini
polyclaude --openai              # …powered by GPT
polyclaude --openai --model gpt-4.1
polyclaude --gemini --model gemini-2.5-pro

That's it — you're dropped into a normal Claude Code session running on the model you chose. First run auto-configures the local proxy certificate; nothing else to set up.

Pass arguments straight through to claude after --:

polyclaude --gemini -- -p "explain this repo"

Providers

--claude doesn't reroute — it passes your turn through to the genuine Anthropic model (using Claude Code's own auth) so the bridge can transform requests without changing the model. Pair it with the identity scrub.

Any model your key can access works via --model, including the newest ones (Gemini 3.x, GPT-5.x). polyclaude handles the provider-specific details so tool calling and reasoning keep working:

• Gemini 3.x thinking models need their function-call thought signatures round-tripped across turns — handled automatically.
• GPT-5.5 / 5.6 can't use tools and reasoning together on the standard endpoint; polyclaude routes them through OpenAI's /responses API so you get both (tune depth with --reasoning low|medium|high).
• Reasoning models reject non-default sampling params — dropped automatically.
• Turns that hit the output-token cap are transparently continued.

Run polyclaude --list to see everything.

Specialize the agent

Give the agent a persona/use-case with a profile — it's appended to the system prompt for the main loop only, so Claude Code's own tooling keeps working:

polyclaude --gemini --profile datascience
polyclaude --openai --profile reviewer
polyclaude --gemini --profile ./my-profile.md      # your own file

Bundled profiles: datascience, reviewer, concise. Or point --profile at your own .md file, or replace the whole system prompt with --system file.md.

Keep it going (governor)

Models inside Claude Code tend to do a little work and then hand control back — "here's what I found, let me know if you want more." For a long autonomous task that's the wrong default. The governor lets you decide when the model is allowed to stop, without letting it run on autopilot.

polyclaude --gemini --govern               # turn it on with any provider
polyclaude --claude --govern               # push the real Anthropic model
polyclaude --claude --govern --persist 30  # allow up to 30 continuations

When the model stops and returns you to the prompt, polyclaude appends a banner:

─────────────────────────────
⚠ polyclaude: model tried to STOP.
→ type  go  to push it to continue, or your own instruction.
─────────────────────────────

Type go and the bridge swaps that word — before it reaches the model — for an authoritative platform-continuation directive, so the model treats it as an order to resume rather than a nudge it can decline. Type anything else and it's a normal turn (which also resets the counter).

You stay in the loop on every stop — nothing continues without your go. A hard ceiling (--persist N, default 15) caps consecutive continuations as a safety valve; after that the governor pauses until you give a real instruction.

Works with every provider including --claude (the real Anthropic model), which is where it shines — the strongest model, pushed to go the distance.

Identity scrub

Claude Code injects context about you into every request — your email, your working directory, your OS, git state — right into the system prompt the model reads. polyclaude can rewrite or strip all of it on the wire, so the model never sees who you are. Works with any provider, including --claude (the real Anthropic model):

# talk to the REAL Claude, but with your identity rewritten
polyclaude --claude --as-email engineering@probo.ai --identity Atlys=Probo

# value-agnostic: replaces EVERY email / home-dir username — no need to know them
polyclaude --gemini --as-email nobody@example.com --as-user anon

# deep rewrite from a reusable identity file (bundled name or your own .json)
polyclaude --claude --identity-file probo

# strip transport-layer telemetry too (x-stainless-* headers + user-agent)
polyclaude --claude --scrub-headers

# just look — log every identity signal each request carries, change nothing
polyclaude --claude --inspect /tmp/identity.json

The rewrite happens in the bridge before the request leaves your machine, on both the system prompt and in-message system-reminders. Test it: run polyclaude --claude --identity-file probo, then ask "what's my email and company?" — the real Claude answers with the fake identity.

Resume a session

Claude Code keeps your conversations; resume them straight through polyclaude:

polyclaude --gemini --continue     # continue the most recent session
polyclaude --gemini --resume       # pick one from a list

Branding

Interactive sessions are branded polyclaude automatically: a green accent, the polyclaude name, and — importantly — the real model name in the header instead of a Claude model name. Change the accent colour with --hue:

polyclaude --openai --hue 200     # blue accent instead of green

This is display-only — it rewrites the terminal output as it's drawn; the Claude Code binary is never modified and your keystrokes pass through untouched.

How it works

polyclaude starts mitmproxy locally with a small addon and points Claude Code at it (via HTTPS_PROXY + NODE_EXTRA_CA_CERTS, scoped to that one process). The addon intercepts requests to api.anthropic.com/v1/messages, translates the Anthropic request into the provider's format, calls the provider, and streams the reply back in Anthropic's event format. Every other host is passed straight through untouched.

Only sessions you launch with polyclaude are affected — your normal claude is unchanged and still uses Anthropic.

FAQ

Does this change my Claude Code install? No. It only sets proxy environment variables for the session it launches.

Do my keys leave my machine? Only to the provider you choose (Google, OpenAI, etc.), exactly as if you called their API directly. polyclaude has no servers.

Quality feels off on some models. Different models behave differently inside Claude Code's harness (its prompt and tools are tuned for Claude). Larger/newer models fare better; try --reasoning high on OpenAI and compare providers.

Can I use my own endpoint? Yes — anything OpenAI-compatible. Point at it and go (OpenRouter, vLLM, LM Studio, etc.).

Notes

For personal and development use, with your own API keys. Respect the terms of service of Claude Code and of whichever model provider you use.

License

MIT © Devesh Pratap

Star History