A multi-provider gateway for Claude Code and other coding agents.
It is an independent implementation. It is not affiliated with Anthropic.
What v0.8 includes
Gateway protocols
- Anthropic Messages:
/v1/messages - Anthropic token counting:
/v1/messages/count_tokens - OpenAI Responses-compatible:
/v1/responses - model discovery:
/v1/models - health:
/health - local Admin UI:
/admin - authenticated Admin API:
/api/admin/*
Agent features
- streaming SSE
- tool definitions and tool calls
- tool results
- images
- reasoning/thinking metadata pass-through
- Claude tier routing: Fable / Opus / Sonnet / Haiku
- no-thinking gateway IDs
- ordered model fallback
- provider health backoff
- per-provider concurrency
- rate-window control
- stable public model identity
Provider layer
The catalog contains the broad provider set used by the project:
NVIDIA NIM, OpenRouter, Groq, OpenAI, xAI, QwenCloud, Together, DeepInfra, SiliconFlow, Nebius, Chutes, Featherless, ZenMux, W&B Inference, Azure OpenAI, Google AI Studio, Google Vertex, DeepSeek, Mistral, Codestral, OpenCode Zen, OpenCode Go, Vercel AI Gateway, Amazon Bedrock, Hugging Face, Cohere, GitHub Models, Wafer, Kimi, Kimi Code, MiniMax, Cerebras, SambaNova, Kilo, Fireworks, Novita, Cloudflare Workers AI, Z.ai, TokenRouter, NaraRoute, Poolside, LLM7, Ollama Cloud, LM Studio, llama.cpp and Ollama.
Provider entries are not fake claims of universal support: providers with unusual authentication/protocols require a dedicated adapter. The common OpenAI-compatible providers use the shared transport.
Coding-agent launcher layer
The architecture has launcher adapters for:
- Claude Code
- Codex
- Pi
- OpenCode
- Cline
- Hermes
- DeepSeek Harness
- Grok Build
- Muse Code
A launcher simply prepares the local proxy environment and delegates arguments to the installed client.
Architecture
Coding Agents / IDEs
|
+-----------+-----------+
| |
Anthropic Messages OpenAI Responses
| |
+-----------+-----------+
|
FastAPI Gateway
|
Model Router
|
+-----------+-----------+
| |
Primary Fallbacks
| |
+-----------+-----------+
|
Provider Runtime
|
+--------------+--------------+
| | |
OpenAI-compatible Specialized Local
adapter adapter runtime
| | |
APIs provider API Ollama/LM Studio
The architecture deliberately separates wire protocols from routing and provider code. This mirrors the important architectural boundary in the current reference project: HTTP adapters, application routing/execution, provider runtime, CLI adapters and optional messaging are separate concerns. citeturn0search0
Install
Python 3.10+.
python -m venv .venv
Windows:
.venv\Scripts\Activate.ps1
python -m pip install -r requirements.txt
copy .env.example .env
macOS/Linux:
source .venv/bin/activate
python -m pip install -r requirements.txt
cp .env.example .env
Start:
python -m my_free_code
Default address:
http://127.0.0.1:8082
Configure a model
Example:
MODEL=open_router/openrouter/free
MODEL_SONNET=deepseek/deepseek-chat
MODEL_HAIKU=groq/llama-3.3-70b-versatile
MODEL_OPUS=nvidia_nim/meta/llama-3.3-70b-instruct
FALLBACK_MODELS=deepseek/deepseek-chat,ollama/llama3.1
Then set the corresponding API keys in .env.
The public model identity stays as the gateway model even when a request is routed to another upstream provider.
Claude Code
$env:ANTHROPIC_BASE_URL="http://127.0.0.1:8082"
$env:ANTHROPIC_AUTH_TOKEN="local"
$env:CLAUDE_CODE_ENABLE_GATEWAY_MODEL_DISCOVERY="1"
claude
Or:
python -m my_free_code.cli.mfc claude
Other clients
The launcher abstraction supports:
python -m my_free_code.cli.mfc codex
python -m my_free_code.cli.mfc pi
python -m my_free_code.cli.mfc opencode
python -m my_free_code.cli.mfc cline
python -m my_free_code.cli.mfc hermes
python -m my_free_code.cli.mfc deepseek-harness
python -m my_free_code.cli.mfc grok
python -m my_free_code.cli.mfc muse
The installed client must already exist on PATH.
Routing and fallback
For:
MODEL_SONNET=deepseek/deepseek-chat
FALLBACK_MODELS=groq/llama-3.3-70b-versatile,ollama/llama3.1
a Sonnet request follows:
Claude Code
|
v
deepseek/deepseek-chat
|
| failure before output
v
groq/llama-3.3-70b-versatile
|
| failure before output
v
ollama/llama3.1
Once a streaming response has committed output, the gateway does not silently switch providers and duplicate the turn.
Reasoning
The gateway accepts Claude-style thinking intent and keeps it separate from provider-specific request translation.
Supported normalized modes:
auto
on
off
and optional effort:
low
medium
high
Provider adapters can map the normalized reasoning policy to their documented upstream fields.
Admin
Open:
http://127.0.0.1:8082/admin
Authenticated JSON endpoints:
GET /api/admin/status
GET /api/admin/models
GET /api/admin/providers
Local models
Ollama:
OLLAMA_BASE_URL=http://127.0.0.1:11434/v1
MODEL=ollama/llama3.1
LM Studio:
LM_STUDIO_BASE_URL=http://127.0.0.1:1234/v1
MODEL=lmstudio/qwen3.5-coder
llama.cpp:
LLAMACPP_BASE_URL=http://127.0.0.1:8080/v1
MODEL=llamacpp/my-model
Security
This is intended for local use.
- keep
HOST=127.0.0.1 - set a non-trivial
PROXY_AUTH_TOKEN - never commit
.env - do not expose Admin endpoints directly to the Internet
- provider credentials remain in environment/configuration and are never sent to another provider
Tests
pytest -q
The repository includes deterministic tests for routing, protocol conversion, auth, reasoning, model catalog and streaming primitives.
Project structure
my-free-code/
├── my_free_code/
│ ├── api/
│ │ ├── routes.py
│ │ └── admin_routes.py
│ ├── cli/
│ │ ├── claude.py
│ │ ├── launchers.py
│ │ └── mfc.py
│ ├── core/
│ │ ├── anthropic.py
│ │ ├── responses.py
│ │ ├── reasoning.py
│ │ ├── failures.py
│ │ ├── streaming.py
│ │ └── model_catalog.py
│ └── providers/
│ ├── catalog.py
│ ├── adapters.py
│ ├── specialized.py
│ └── runtime.py
├── tests/
├── ARCHITECTURE.md
├── .env.example
├── requirements.txt
├── pyproject.toml
└── README.md
Comments