backdoor

██████╗  █████╗  ██████╗██╗  ██╗██████╗  ██████╗  ██████╗ ██████╗
██╔══██╗██╔══██╗██╔════╝██║ ██╔╝██╔══██╗██╔═══██╗██╔═══██╗██╔══██╗
██████╔╝███████║██║     █████╔╝ ██║  ██║██║   ██║██║   ██║██████╔╝
██╔══██╗██╔══██║██║     ██╔═██╗ ██║  ██║██║   ██║██║   ██║██╔══██╗
██████╔╝██║  ██║╚██████╗██║  ██╗██████╔╝╚██████╔╝╚██████╔╝██║  ██║
╚═════╝ ╚═╝  ╚═╝ ╚═════╝╚═╝  ╚═╝╚═════╝  ╚═════╝  ╚═════╝ ╚═╝  ╚═╝

The problem

Claude Code is genuinely in a league of its own as a coding agent. The tool use, the agentic loops, the way it navigates a codebase — nothing else comes close.

But it's hardwired to Anthropic's API. You pay Anthropic's prices. You use Anthropic's models. You play by Anthropic's rules.

That's a lot of lock-in for a tool that could theoretically run on anything.

The fix

Backdoor is a lightweight proxy that intercepts Claude Code's API calls and reroutes them to any OpenAI-compatible provider. It translates the request format, streams the response back, handles tool use — everything. Claude Code has no idea anything changed.

Before:   Claude Code  →  Anthropic API  →  $$$
After:    Claude Code  →  Backdoor  →  literally anything

Three commands to set up. One line to switch providers.

The numbers don't lie

We ran 500 million tokens through Claude Code. Same workload. Two different backends.

That's not a typo. $9,834 vs $5.34. Nearly 2,000x cheaper at scale.

If you're an individual developer, that's the difference between a credit card bill that hurts and one that rounds to zero. If you're a team running Claude Code across multiple engineers, this is the difference between a budget line item that gets cut and one nobody notices.

The harness is free. The intelligence is cheap. That's the whole point.

What you actually get

🆓 Run it free. NVIDIA NIM and Groq both have free tiers with thousands of requests per month. Backdoor works with both out of the box.

💸 Run it cheap. DeepSeek costs roughly 95% less than Claude's API. If you're using Claude Code daily, that's the difference between a coffee and a car payment.

🔒 Run it private. Point Backdoor at a local Ollama instance and nothing ever leaves your machine. No API. No logs. No data sent anywhere.

🔀 Run whatever model you want. One config change and you're on a completely different AI. Benchmark Llama vs. Mistral vs. DeepSeek against your actual work, using the best coding agent UI ever built as the harness.

🌍 200+ models, one tool. Wire up OpenRouter and you have access to every major model on earth — Gemini, Grok, Qwen, Command R, all of them — through a single API key.

Get running in 60 seconds

git clone https://github.com/ajsai47/backdoor
cd backdoor
./backdoor   # animated setup wizard — picks your provider, writes your config, launches Claude Code

That's it. The wizard handles everything: animated intro, provider selection, API key entry, and launching your first session. If you'd rather skip the wizard and configure manually, copy .env.example to .env, fill it in, and run ./run.sh directly.

Pick your provider

Edit three lines in .env and you're on a different AI:

PROVIDER_BASE_URL=https://api.deepseek.com/v1
PROVIDER_API_KEY=your-key
PROVIDER_MODEL=deepseek-chat

Why this exists

I'm a huge fan of Anthropic. Genuinely. Their models are some of the most impressive technology I've ever used, and Claude Code is the best coding agent ever built — it's not close.

This project isn't a knock on them. It's a tribute to how good the harness is.

What I believe is simple: the best tools should be accessible to everyone, not just the people who can afford the top-tier API. A student, a solo developer, a builder in a country where $9,000/month in API costs is unthinkable — they deserve to experience what Claude Code can do. Vendor lock-in shouldn't be the thing standing between a great developer and a great tool.

Anthropic built the magic. Backdoor makes sure everyone gets to use it.

Run Claude Code. Bring your own model. Keep building.

Open source. No strings.

MIT licensed. Read the code — it's clean, it's simple, it's less than 600 lines. Fork it, change it, build on it.

If a new AI provider launches tomorrow, you can use it with Backdoor the same day. No waiting for a pull request. Just drop in the URL.

Bonus: control Claude Code from your phone

Set TELEGRAM_BOT_TOKEN and TELEGRAM_ALLOWED_USER_ID in .env and you can trigger Claude Code sessions from Telegram. Send a prompt from your phone, get the output back in chat.

Troubleshooting

Proxy failed to start
Something is already using port 8082. Either stop the other process or change PORT=8083 in your .env.

PROVIDER_API_KEY is not set
Open .env and replace your-api-key-here with your actual key.

Claude Code is not installed
Install it from claude.ai/code.

Model not found or 404 from the provider
The model name in PROVIDER_MODEL doesn't match what the provider expects. Check the provider's docs for the exact model slug — they vary (e.g. DeepSeek uses deepseek-chat, Groq uses llama-3.3-70b-versatile).

Responses feel slow
Switch to Groq — it's the fastest inference available and has a free tier.

Nothing is logged
Check proxy.log in the project directory. If it's empty, the proxy didn't start at all — run uv run uvicorn server:app directly to see the error.

How it works under the hood

Claude Code talks to localhost:8082 thinking it's Anthropic. Backdoor receives the request, translates it from Anthropic's Messages API format to OpenAI's chat completions format, forwards it to your chosen provider, and streams the response back — translated back into Anthropic's SSE format in real time. Tool calls, streaming deltas, token counts — all handled transparently.

A handful of Claude Code's internal housekeeping requests (quota probes, title generation, etc.) are intercepted and short-circuited locally so they don't burn your provider quota.