Bring Your Own Model — AI skill for Claude Code

bring-your-own-model — run Claude Code on any model

[](https://github.com/askalf/bring-your-own-model/actions/workflows/ci.yml) [](https://scorecard.dev/viewer/?uri=github.com/askalf/bring-your-own-model)

**Archived — August 2026.** This bridge works as documented, and the code stays up to read. But it is a finished experiment, not a maintained product: no further releases are planned. The routing layers we do maintain are [dario](https://github.com/askalf/dario) (your Claude subscription behind one local endpoint) and [hybrid](https://github.com/askalf/hybrid) (local-first model routing).

Claude Code speaks one wire protocol: the Anthropic Messages API. This is a tiny local proxy that translates that protocol **to and from OpenAI** (the Responses API and Chat Completions), so you can point Claude Code at GPT-5.6 Sol, gpt-4o, or any OpenAI-compatible endpoint — and drive the full agentic harness with a different model underneath.

Tools, streaming, reasoning, the whole loop. Claude Code never knows.

Zero runtime dependencies — just Node's built-in `http` and two pure translator modules. The whole thing is a few hundred lines you can read.

The demo is GPT-5.6 Sol reading a failing test, reasoning about *why* it fails, and fixing its own code — live inside Claude Code, over the OpenAI wire:

**[▶ Watch the demo](demo/cc-sol-live-cap.mp4)** — `demo/cc-sol-live-cap.mp4`

Quickstart

npm install
npm run build

# your OpenAI key (or any OpenAI-compatible provider's)
export OPENAI_API_KEY=sk-...

# start the bridge, forcing every request to gpt-5.6-sol
node dist/cli.js --model gpt-5.6-sol --port 8788

Then point Claude Code at it — in a separate shell:

export ANTHROPIC_BASE_URL=http://localhost:8788
export ANTHROPIC_API_KEY=any-placeholder
claud