Quick Start

1. Install/Update The Proxy

macOS/Linux:

curl -fsSL "https://github.com/Alishahryar1/free-claude-code/blob/main/scripts/install.sh?raw=1" | sh

Windows PowerShell:

irm "https://github.com/Alishahryar1/free-claude-code/blob/main/scripts/install.ps1?raw=1" | iex

Review the installers at scripts/install.sh and scripts/install.ps1. They install Claude Code and Codex when missing, then install or update the proxy. Re-run these commands to update to the latest version.

To remove only Free Claude Code (not uv, Claude Code, Codex, or the uv-managed Python runtime):

macOS/Linux:

curl -fsSL "https://raw.githubusercontent.com/Alishahryar1/free-claude-code/main/scripts/uninstall.sh" | sh

Windows PowerShell:

irm "https://raw.githubusercontent.com/Alishahryar1/free-claude-code/main/scripts/uninstall.ps1" | iex

Review scripts/uninstall.sh and scripts/uninstall.ps1. They remove the FCC uv tool and always delete ~/.fcc/. Stop any running fcc-server, fcc-claude, fcc-codex, fcc-init, or free-claude-code process before uninstalling.

2. Start The Proxy

fcc-server

After startup, Uvicorn prints the proxy bind address and the app logs the admin URL:

INFO:     Admin UI: http://127.0.0.1:8082/admin (local-only)

Many terminals make these clickable. Use your configured PORT if it is not 8082.

3. Open The Admin UI And Configure NVIDIA NIM

Open the Admin UI URL from the terminal output.

Need an NVIDIA NIM API key? Use the NVIDIA NIM provider section below, then scroll back up here.

Paste your NVIDIA NIM API key into NVIDIA_NIM_API_KEY, then click Validate and Apply.

The default model is already set to nvidia_nim/nvidia/nemotron-3-super-120b-a12b. You can change it later from the same Admin UI.

4. Run Your Coding Agent

Keep fcc-server running while you work.

Claude Code

fcc-claude

fcc-claude reads the current configured port and auth token each time it starts, sets the Claude Code environment variables (including a 190k-token CLAUDE_CODE_AUTO_COMPACT_WINDOW for auto-compaction), and then launches the real claude command. When proxy auth is disabled, it still passes ANTHROPIC_AUTH_TOKEN=fcc-no-auth so newer Claude Code versions do not stop at their local login gate before contacting the proxy.

Codex

fcc-codex

fcc-codex reads the same port and auth token, registers an ephemeral fcc model provider that points at the local proxy's /v1/responses endpoint, generates a Codex model catalog from the proxy's /v1/models response, sets FCC_CODEX_API_KEY from the Admin UI auth token, strips official OPENAI_* credentials from the child environment, and then launches the real codex command. Type /model inside Codex to open its native picker. Pass through Codex args as usual, for example fcc-codex exec "hello".

Choose A Provider

Pick one provider, enter its key or local URL in the Admin UI, and set MODEL to a provider-prefixed model slug. MODEL is the fallback. MODEL_OPUS, MODEL_SONNET, and MODEL_HAIKU can override routing for Claude Code's model tiers.

1. NVIDIA NIM

Get a key at build.nvidia.com/settings/api-keys.

In the Admin UI, paste it into NVIDIA_NIM_API_KEY. The default MODEL is nvidia_nim/nvidia/nemotron-3-super-120b-a12b.

Popular examples:

• nvidia_nim/nvidia/nemotron-3-super-120b-a12b
• nvidia_nim/z-ai/glm5.1
• nvidia_nim/moonshotai/kimi-k2.5
• nvidia_nim/minimaxai/minimax-m2.5

Browse models at build.nvidia.com.

2. OpenRouter

Get a key at openrouter.ai/keys.

In the Admin UI, paste it into OPENROUTER_API_KEY, then set MODEL to an OpenRouter slug such as open_router/openrouter/free.

Browse all models or free models.

3. Google AI Studio (Gemini)

Get a Gemini API key at Google AI Studio (see Google's Gemini OpenAI compatibility docs).

In the Admin UI, paste it into GEMINI_API_KEY, then set MODEL to a Gemini model slug such as gemini/models/gemini-3.1-flash-lite.

The Gemini API exposes an OpenAI-compatible endpoint at https://generativelanguage.googleapis.com/v1beta/openai/. Free tier quotas are per-model; prompts may be used to improve Google's products outside the UK/CH/EEA/EU unless your account region says otherwise—see Google's terms.

Popular examples:

• gemini/models/gemini-3.1-flash-lite

4. DeepSeek

Get a key at platform.deepseek.com/api_keys.

In the Admin UI, paste it into DEEPSEEK_API_KEY, then set MODEL to a DeepSeek slug such as deepseek/deepseek-chat.

This provider uses DeepSeek's Anthropic-compatible endpoint, not the OpenAI chat-completions endpoint.

5. Mistral La Plateforme

Mistral hosts an OpenAI-compatible Chat Completions API at https://api.mistral.ai/v1. Activate the Experiment plan on console.mistral.ai for free-tier API access with rate limits (upgrade for higher quotas).

In the Admin UI, paste your API key into MISTRAL_API_KEY, then set MODEL to a Mistral model slug such as mistral/devstral-small-latest or mistral/mistral-small-latest.

Popular examples:

• mistral/devstral-small-latest
• mistral/mistral-small-latest

Browse models at Mistral documentation.

6. Mistral Codestral

Mistral's Codestral gateway uses a separate API key from La Plateforme: provision CODESTRAL_API_KEY, then route with the mistral_codestral/ prefix. The default upstream is https://codestral.mistral.ai/v1 (OpenAI-compatible Chat Completions; same request shaping as the mistral provider). See Mistral's coding / FIM domains; the curated free LLM API list summarizes typical Codestral access terms.

Popular examples:

• mistral_codestral/codestral-latest

7. OpenCode Zen

Get an API key at opencode.ai/auth.

In the Admin UI, paste it into OPENCODE_API_KEY, then set MODEL to an OpenCode Zen model slug such as opencode/gpt-5.3-codex. The same OPENCODE_API_KEY powers OpenCode Go (below); use opencode_go/ slugs there.

OpenCode Zen is a curated model gateway that provides access to models from Anthropic, OpenAI, Google, DeepSeek, and more through a single API key and OpenAI-compatible endpoint at https://opencode.ai/zen/v1.

Popular examples:

• opencode/gpt-5.3-codex
• opencode/claude-sonnet-4
• opencode/deepseek-v4-flash-free (free)
• opencode/gemini-3-flash
• opencode/big-pickle (free)
• opencode/glm-5.1

Browse available models at opencode.ai.

8. OpenCode Go

Get an API key at opencode.ai/auth (same as OpenCode Zen).

In the Admin UI, use OPENCODE_API_KEY, then set MODEL to an OpenCode Go model slug such as opencode_go/minimax-m2.7.

OpenCode Go is a subscription gateway with its own curated catalog and OpenAI-compatible endpoint at https://opencode.ai/zen/go/v1. It shares the same OpenCode API key as Zen; only the slug prefix (opencode_go/ vs opencode/) and upstream path differ.

Popular examples:

• opencode_go/minimax-m2.7

Browse available models at opencode.ai.

9. Wafer

Get a key from wafer.ai. In the Admin UI, paste it into WAFER_API_KEY, then set MODEL to a Wafer Pass model such as wafer/DeepSeek-V4-Pro.

Popular examples:

• wafer/DeepSeek-V4-Pro
• wafer/MiniMax-M2.7
• wafer/Qwen3.5-397B-A17B
• wafer/GLM-5.1

This provider uses Wafer's Anthropic-compatible endpoint at https://pass.wafer.ai/v1/messages.

10. Kimi

Get a key at platform.moonshot.ai/console/api-keys.

In the Admin UI, paste it into KIMI_API_KEY, then set MODEL to a Kimi slug such as kimi/kimi-k2.5.

This provider calls Kimi's Anthropic-compatible Messages API (https://api.moonshot.ai/anthropic/v1/messages; model discovery uses OpenAI-compat GET https://api.moonshot.ai/v1/models). It is not the OpenAI Chat Completions path.

Browse models at platform.moonshot.ai.

11. Cerebras Inference

Sign up and create an API key in the Cerebras Cloud Console (see Quickstart).

In the Admin UI, set CEREBRAS_API_KEY, then route with MODEL such as cerebras/llama3.1-8b or cerebras/gpt-oss-120b (ids from List models).

Cerebras exposes an OpenAI-compatible API at https://api.cerebras.ai/v1 (OpenAI compatibility). Non-standard request fields should go in extra_body when using the OpenAI client; see the same page. For reasoning models and parameters, see Reasoning. This proxy follows other OpenAI-compat adapters for thinking via reasoning_content when Claude-style thinking is enabled.

12. Groq

Get an API key at console.groq.com/keys.

In the Admin UI, paste it into GROQ_API_KEY, then set MODEL to a Groq OpenAI-compat model slug such as groq/llama-3.3-70b-versatile.

Groq routes through https://api.groq.com/openai/v1 (OpenAI-compatible Chat Completions). Some request fields yield HTTP 400; this adapter strips known-unsupported shapes (documented in Groq's compatibility notes).

Reasoning-heavy models expose extra knobs documented under Groq reasoning. This release mirrors other OpenAI-compat adapters for thinking via reasoning_content deltas when Claude-style thinking is enabled; you can tune advanced parameters through request extra_body when needed.

Browse models at console.groq.com/docs/models.

13. Fireworks AI

Get an API key at fireworks.ai/account/api-keys.

In the Admin UI, paste it into FIREWORKS_API_KEY, then set MODEL to a Fireworks model slug such as fireworks/accounts/fireworks/models/llama-v3p3-70b-instruct.

Fireworks exposes an Anthropic-compatible Messages API at https://api.fireworks.ai/inference/v1/messages (same inference host as before; Chat Completions is not used here). Vendor-specific JSON keys can still be merged from request extra_body when allowed.

Browse models at fireworks.ai/models.

14. Z.ai

Get an API key at Z.ai/manage-apikey/apikey-list.

In the Admin UI, paste it into ZAI_API_KEY, then set MODEL to a Z.ai model slug such as zai/glm-5.1.

This provider calls Z.ai's Anthropic-compatible Messages API (https://api.z.ai/api/anthropic/v1/messages). The former OpenAI Coding Plan base (https://api.z.ai/api/coding/paas/v4) is not used by this gateway.

Popular examples:

• zai/glm-5.1
• zai/glm-5-turbo

Browse models at Z.ai.

15. LM Studio

Start LM Studio's local server and load a model. In the Admin UI, keep or update LM_STUDIO_BASE_URL, then set MODEL to the model identifier shown by LM Studio, prefixed with lmstudio/.

Prefer models with tool-use support for Claude Code workflows.

16. llama.cpp

Start llama-server with an Anthropic-compatible /v1/messages endpoint and enough context for Claude Code requests.

In the Admin UI, keep or update LLAMACPP_BASE_URL, then set MODEL to the local model slug, prefixed with llamacpp/.

For local coding models, context size matters. If llama.cpp returns HTTP 400 for normal Claude Code requests, increase --ctx-size and verify the model/server build supports the requested features.

17. Ollama

Run Ollama and pull a model:

ollama pull llama3.1
ollama serve

In the Admin UI, keep or update OLLAMA_BASE_URL, then set MODEL to the same tag shown by ollama list, prefixed with ollama/.

OLLAMA_BASE_URL is the Ollama server root; do not append /v1. Example model slugs include ollama/llama3.1 and ollama/llama3.1:8b.

18. Mix Providers By Model Tier

Each model tier can use a different provider by setting MODEL_OPUS, MODEL_SONNET, and MODEL_HAIKU in the Admin UI. Leave a tier blank to inherit MODEL. These tier overrides apply to Claude model names that contain opus, sonnet, or haiku. Codex uses the Admin MODEL default through fcc-codex unless a session requests a provider-prefixed slug directly.

For example, you can route Opus to nvidia_nim/moonshotai/kimi-k2.6, Sonnet to open_router/openrouter/free, Haiku to lmstudio/qwen3.5-coder, and keep the fallback MODEL on zai/glm-5.1.

Connect Your Client

1. Claude Code CLI

For terminal use, prefer the installed launcher:

fcc-claude

The Admin UI manages proxy config, restarts the server when runtime settings change, and fcc-claude reads the current Admin UI-managed port and auth token every time it starts. It also sets CLAUDE_CODE_AUTO_COMPACT_WINDOW to 190000 for auto-compaction. When proxy auth is blank, fcc-claude injects ANTHROPIC_AUTH_TOKEN=fcc-no-auth only to satisfy Claude Code's local login check; the proxy still treats blank auth as disabled.

2. Codex CLI

For terminal use, prefer the installed launcher:

fcc-codex

The installer provisions Codex when it is missing (npm install -g @openai/codex). fcc-codex injects ephemeral Codex config on every launch:

• model_provider=fcc
• model_providers.fcc.base_url=http://127.0.0.1:<PORT>/v1
• model_providers.fcc.env_key=FCC_CODEX_API_KEY
• model_providers.fcc.wire_api=responses
• model_catalog_json=~/.fcc/codex-model-catalog.json

The Admin UI auth token is reused as FCC_CODEX_API_KEY. Official OpenAI credentials are stripped from the child environment so traffic stays on the local proxy. The generated model catalog lets Codex's native /model picker list provider-selectable FCC model slugs. If the catalog cannot be fetched or written, fcc-codex warns and still launches without picker injection.

Advanced manual setup

If you launch codex directly, point it at the proxy with equivalent config:

codex \
  -c 'model_provider="fcc"' \
  -c 'model_providers.fcc.name="Free Claude Code"' \
  -c 'model_providers.fcc.base_url="http://127.0.0.1:8082/v1"' \
  -c 'model_providers.fcc.env_key="FCC_CODEX_API_KEY"' \
  -c 'model_providers.fcc.wire_api="responses"' \
  exec "hello"

Set FCC_CODEX_API_KEY to the same value as ANTHROPIC_AUTH_TOKEN in the Admin UI.

3. Claude Code in VS Code

Install the Claude Code extension. Open Settings, search for claude-code.environmentVariables, choose Edit in settings.json, and add:

"claudeCode.environmentVariables": [
  { "name": "ANTHROPIC_BASE_URL", "value": "http://localhost:8082" },
  { "name": "ANTHROPIC_AUTH_TOKEN", "value": "freecc" },
  { "name": "CLAUDE_CODE_ENABLE_GATEWAY_MODEL_DISCOVERY", "value": "1" },
  { "name": "CLAUDE_CODE_AUTO_COMPACT_WINDOW", "value": "190000" }
]

Reload the extension. If the extension shows a login screen, choose the Anthropic Console path once; the local proxy still handles model traffic after the environment variables are active.

4. Codex in VS Code

Install the Codex extension. The extension shares the same user-level Codex config as the CLI (~/.codex/config.toml on macOS/Linux, %USERPROFILE%\.codex\config.toml on Windows).

Create or edit that file with the fcc provider pointing at your local proxy:

model_provider = "fcc"
model = "nvidia_nim/nvidia/nemotron-3-super-120b-a12b"

[model_providers.fcc]
name = "Free Claude Code"
base_url = "http://127.0.0.1:8082/v1"
env_key = "FCC_CODEX_API_KEY"
wire_api = "responses"

Set model to your Admin UI MODEL value. Replace 8082 if your proxy uses a different PORT.

Store the proxy auth token in ~/.codex/auth.json (or the Windows equivalent):

{
  "FCC_CODEX_API_KEY": "freecc"
}

Use the same value as ANTHROPIC_AUTH_TOKEN in the Admin UI. Restart VS Code after changing these files. On Windows with WSL-backed Codex, edit the WSL ~/.codex/ files instead and enable chatgpt.runCodexInWindowsSubsystemForLinux in VS Code settings when needed.

5. Claude Code in JetBrains ACP

Edit the installed Claude ACP config:

• Windows: C:\Users\%USERNAME%\AppData\Roaming\JetBrains\acp-agents\installed.json
• Linux/macOS: ~/.jetbrains/acp.json

Set the environment for acp.registry.claude-acp:

"env": {
  "ANTHROPIC_BASE_URL": "http://localhost:8082",
  "ANTHROPIC_AUTH_TOKEN": "freecc",
  "CLAUDE_CODE_ENABLE_GATEWAY_MODEL_DISCOVERY": "1",
  "CLAUDE_CODE_AUTO_COMPACT_WINDOW": "190000"
}

Restart the IDE after changing the file.

Optional Integrations

For every integration below, change managed proxy settings only in the Admin UI at /admin: edit fields, click Validate, then Apply. The footer shows where the managed config is stored; this README does not walk through editing that file by hand.

1. Discord And Telegram Bots

The bot wrapper runs Claude Code sessions remotely, streams progress, supports reply-based conversation branches, and can stop or clear tasks. Discord and Telegram bots use Claude Code today; use fcc-codex or the Codex VS Code extension for Codex sessions.

Discord

1. Create the bot in the Discord Developer Portal.
2. Enable Message Content Intent.
3. Invite the bot with read, send, and message history permissions.
4. Copy the bot token and the numeric channel ID (or IDs) where the bot should respond.

Telegram

1. Create a bot with @BotFather and copy the bot token.
2. Get your numeric user ID from @userinfobot so only you can use the bot.

Configure in the Admin UI

1. With fcc-server running, open the Admin UI URL from the terminal output.
2. In the sidebar, choose Messaging.
3. Set Messaging Platform to discord or telegram.
4. For Discord, paste Discord Bot Token and Allowed Discord Channels. For Telegram, paste Telegram Bot Token and Allowed Telegram User ID.
5. Set Allowed Directory to an absolute path on the machine running the proxy—the workspace root the bot may use.
6. Click Validate, then Apply. Restart the server if the UI says one is required.

Admin UI → Messaging (platform, bots, and Voice)

Useful commands

• /stop cancels a task; reply to a task message to stop only that branch.
• /clear resets sessions; reply to clear one branch.
• /stats shows session state.

2. Voice Notes

Voice notes work on Discord and Telegram after you extend your Free Claude Code install with the matching optional extras.

macOS/Linux:

# NVIDIA NIM transcription (Riva gRPC)
curl -fsSL "https://github.com/Alishahryar1/free-claude-code/blob/main/scripts/install.sh?raw=1" | sh -s -- --voice-nim

# Local Whisper (CPU or CUDA)
curl -fsSL "https://github.com/Alishahryar1/free-claude-code/blob/main/scripts/install.sh?raw=1" | sh -s -- --voice-local

# Both backends
curl -fsSL "https://github.com/Alishahryar1/free-claude-code/blob/main/scripts/install.sh?raw=1" | sh -s -- --voice-all

# Local Whisper with CUDA
curl -fsSL "https://github.com/Alishahryar1/free-claude-code/blob/main/scripts/install.sh?raw=1" | sh -s -- --voice-local --torch-backend cu130

Windows PowerShell:

# NVIDIA NIM transcription (Riva gRPC)
& ([scriptblock]::Create((irm "https://github.com/Alishahryar1/free-claude-code/blob/main/scripts/install.ps1?raw=1"))) -VoiceNim

# Local Whisper (CPU or CUDA)
& ([scriptblock]::Create((irm "https://github.com/Alishahryar1/free-claude-code/blob/main/scripts/install.ps1?raw=1"))) -VoiceLocal

# Both backends
& ([scriptblock]::Create((irm "https://github.com/Alishahryar1/free-claude-code/blob/main/scripts/install.ps1?raw=1"))) -VoiceAll

# Local Whisper with CUDA
& ([scriptblock]::Create((irm "https://github.com/Alishahryar1/free-claude-code/blob/main/scripts/install.ps1?raw=1"))) -VoiceLocal -TorchBackend cu130

Restart fcc-server after reinstalling.

In the Admin UI, open Messaging and scroll to Voice. Turn on Voice Notes, choose Whisper Device (cpu, cuda, or nvidia_nim), set Whisper Model, and enter Hugging Face Token when your setup needs it. For nvidia_nim transcription, install the voice extra and set NVIDIA NIM API Key on the Providers view. The screenshot above shows the Voice block in the same view.

How It Works

Diagram source: assets/how-it-works.mmd.

Important pieces:

• FastAPI exposes Anthropic-compatible routes such as /v1/messages, /v1/messages/count_tokens, and /v1/models, plus OpenAI Responses at /v1/responses.
• Claude Code sends Anthropic Messages; Codex sends OpenAI Responses SSE to the same proxy.
• Responses requests convert to Anthropic Messages internally, then share the same model router, normalizer, and provider adapters.
• fcc-codex registers a custom fcc provider that points Codex at the local proxy's /v1/responses endpoint.
• Model routing resolves Claude model names to MODEL_OPUS, MODEL_SONNET, MODEL_HAIKU, or MODEL.
• NIM, OpenCode Zen, and OpenCode Go use OpenAI chat streaming translated into Anthropic SSE.
• Wafer, OpenRouter, DeepSeek, Kimi, Fireworks AI, Z.ai, LM Studio, llama.cpp, and Ollama use Anthropic Messages style transports where applicable (with provider-specific quirks and model-list URLs).
• The proxy normalizes thinking blocks, tool calls, token usage metadata, and provider errors into the shape each client expects.
• Request optimizations answer trivial Claude Code probes locally to save latency and quota.

Development

1. Project Structure

free-claude-code/
├── server.py              # ASGI entry point
├── api/                   # FastAPI routes, service layer, routing, optimizations
├── core/                  # Shared Anthropic protocol helpers, SSE, OpenAI Responses
│   └── openai_responses/  # Responses ↔ Anthropic conversion and SSE mapping
├── providers/             # Provider transports, registry, rate limiting
├── messaging/             # Discord/Telegram adapters, sessions, voice
├── cli/                   # Package entry points and client CLI process management
├── config/                # Settings, provider catalog, logging
└── tests/                 # Unit and contract tests

2. Run From Source

Use this path if you are developing or want to run directly from a checkout:

git clone https://github.com/Alishahryar1/free-claude-code.git
cd free-claude-code
uv run uvicorn server:app --host 0.0.0.0 --port 8082

3. Commands

Run the local CI sequence (requires uv on PATH). The local scripts format Python files and apply autofixable Ruff lint fixes before type checking and tests:

./scripts/ci.sh

.\scripts\ci.ps1

Useful flags: --only pytest, --skip pytest, --dry-run (PowerShell: -Only pytest, -Skip pytest, -DryRun).

Or run individual repair/check commands manually:

uv run ruff format
uv run ruff check --fix
uv run ty check
uv run pytest -v --tb=short

GitHub CI remains check-only for Ruff with uv run ruff format --check and uv run ruff check, so required status checks verify committed code.

CI also enforces a ban on # type: ignore / # ty: ignore suppressions; scripts/ci.sh and scripts/ci.ps1 run that grep too.

4. Package Scripts

pyproject.toml installs:

• fcc-server: starts the proxy with configured host and port.
• fcc-init: optional advanced scaffold for ~/.fcc/.env; prefer the Admin UI for normal configuration.
• fcc-claude: launches Claude Code with the configured local proxy URL, an auth-token env var or fcc-no-auth sentinel, model discovery flag, and a 190k CLAUDE_CODE_AUTO_COMPACT_WINDOW for auto-compaction.
• fcc-codex: launches Codex with ephemeral fcc provider config pointing at the local proxy's /v1/responses endpoint, a generated native /model picker catalog, and FCC_CODEX_API_KEY from the Admin UI auth token.
• free-claude-code: compatibility alias for fcc-server.

5. Extending

• Add OpenAI-compatible providers by extending OpenAIChatTransport.
• Add Anthropic Messages providers by extending AnthropicMessagesTransport.
• Extend OpenAI Responses conversion in core/openai_responses/ when Codex adds new request or stream shapes.
• Register provider metadata in config.provider_catalog and factory wiring in providers.registry.
• Add messaging platforms by implementing the MessagingPlatform interface in messaging/.

Contributing

• .env.example lists env key names as a read-only reference for contributors; use the Admin UI to change managed proxy settings.
• Report bugs and feature requests in Issues. For bug always include all model mapping, current model when issue occured and the issue string
• Keep changes small and covered by focused tests.
• Do not open Docker integration PRs.
• Do not open README change PRs just open an issue for it.
• Run the full check sequence before opening a pull request.
• The syntax except X, Y is brought back in python 3.14 final version (not in 3.14 alpha). Keep in mind before opening PRs.

License

MIT License. See LICENSE for details.