Data

Tools

Indexes

Applications

Experiments

Awesome

An open API service indexing awesome lists of open source software.

https://github.com/starbaser/ccproxy

Build mods for Claude Code: Hook any request, modify any response, /model "with-your-custom-model", intelligent model routing using your logic or ours
https://github.com/starbaser/ccproxy

ai ai-gateway ai-proxy ai-tools anthropic claude claude-ai claude-api claude-code claude-max claudecode gemini gemini-cli litellm llm llm-gateway llm-proxy llmops openai openrouter

Last synced: 24 days ago
JSON representation

Build mods for Claude Code: Hook any request, modify any response, /model "with-your-custom-model", intelligent model routing using your logic or ours

Awesome Lists containing this project

README 

          
# `ccproxy` - Claude Code Proxy [![Version](https://img.shields.io/badge/version-1.2.0-blue.svg)](https://github.com/starbased-co/ccproxy)



> [Discord](https://starbased.net/discord)



`ccproxy` unlocks the full potential of your Claude Code by enabling Claude use alongside other LLM providers like OpenAI, Gemini, and Perplexity



It works by intercepting Claude Code's requests through a [LiteLLM Proxy Server](https://docs.litellm.ai/docs/simple_proxy), allowing you to route different types of requests to the most suitable model - keep your unlimited Claude for standard coding, send large contexts to Gemini's 2M token window, route web searches to Perplexity, all while Claude Code thinks it's talking to the standard API.



> 🚀 **v2.0 prerelease is now available** — a ground-up rewrite that drops the LiteLLM proxy server entirely. v2.0 jails your process in a rootless WireGuard namespace, intercepts at the network layer with full TLS inspection, and routes traffic through a DAG-driven hook pipeline. Any LLM client works — not just Claude Code. **[Check it out →](https://github.com/starbased-co/ccproxy/tree/dev)**



## Installation



**Important:** ccproxy must be installed with LiteLLM in the same environment so that LiteLLM can import the ccproxy handler.



### Recommended: Install as uv tool



```bash

# Install from PyPI

uv tool install claude-ccproxy --with 'litellm[proxy]'



# Or install from GitHub (latest)

uv tool install git+https://github.com/starbased-co/ccproxy.git --with 'litellm[proxy]'

```



This installs:



- `ccproxy` command (for managing the proxy)

- `litellm` bundled in the same environment (so it can import ccproxy's handler)



### Alternative: Install with pip



```bash

# Install both packages in the same virtual environment

pip install git+https://github.com/starbased-co/ccproxy.git

pip install 'litellm[proxy]'

```



**Note:** With pip, both packages must be in the same virtual environment.



### Verify Installation



```bash

ccproxy --help

# Should show ccproxy commands



which litellm

# Should point to litellm in ccproxy's environment

```



## Usage



Run the automated setup:



```bash

# This will create all necessary configuration files in ~/.ccproxy

ccproxy install



tree ~/.ccproxy

# ~/.ccproxy

# ├── ccproxy.yaml

# └── config.yaml



# ccproxy.py is auto-generated when you start the proxy



# Start the proxy server

ccproxy start --detach



# Start Claude Code

ccproxy run claude

# Or add to your .zshrc/.bashrc

export ANTHROPIC_BASE_URL="http://localhost:4000"

# Or use an alias

alias claude-proxy='ANTHROPIC_BASE_URL="http://localhost:4000" claude'

```



Congrats, you have installed `ccproxy`! The installed configuration files are intended to be a simple demonstration, thus continuing on to the next section to configure `ccproxy` is **recommended**.



### Configuration



#### `ccproxy.yaml`



This file controls how `ccproxy` hooks into your Claude Code requests and how to route them to different LLM models based on rules. Here you specify rules, their evaluation order, and criteria like token count, model type, or tool usage.



```yaml

ccproxy:

  debug: true



  # OAuth token sources - map provider names to shell commands

  # Tokens are loaded at startup for SDK/API access outside Claude Code

  oat_sources:

    anthropic: "jq -r '.claudeAiOauth.accessToken' ~/.claude/.credentials.json"

    # Extended format with custom User-Agent:

    # gemini:

    #   command: "jq -r '.token' ~/.gemini/creds.json"

    #   user_agent: "MyApp/1.0"



  hooks:

    - ccproxy.hooks.rule_evaluator # evaluates rules against request (needed for routing)

    - ccproxy.hooks.model_router # routes to appropriate model

    - ccproxy.hooks.forward_oauth # forwards OAuth token to provider

    - ccproxy.hooks.extract_session_id # extracts session ID for LangFuse tracking

    # - ccproxy.hooks.capture_headers  # logs HTTP headers (with redaction)

    # - ccproxy.hooks.forward_apikey   # forwards x-api-key header

  rules:

    # example rules

    - name: token_count

      rule: ccproxy.rules.TokenCountRule

      params:

        - threshold: 60000

    - name: web_search

      rule: ccproxy.rules.MatchToolRule

      params:

        - tool_name: WebSearch

    # basic rules

    - name: background

      rule: ccproxy.rules.MatchModelRule

      params:

        - model_name: claude-3-5-haiku-20241022

    - name: think

      rule: ccproxy.rules.ThinkingRule



litellm:

  host: 127.0.0.1

  port: 4000

  num_workers: 4

  debug: true

  detailed_debug: true

```



When `ccproxy` receives a request from Claude Code, the `rule_evaluator` hook labels the request with the first matching rule:



1. `MatchModelRule`: A request with `model: claude-3-5-haiku-20241022` is labeled: `background`

2. `ThinkingRule`: A request with `thinking: {enabled: true}` is labeled: `think`



If a request doesn't match any rule, it receives the `default` label.



#### `config.yaml`



[LiteLLM's proxy configuration file](https://docs.litellm.ai/docs/proxy/config_settings) is where your model deployments are defined. The `model_router` hook takes advantage of [LiteLLM's model alias feature](https://docs.litellm.ai/docs/completion/model_alias) to dynamically rewrite the model field in requests based on rule criteria before LiteLLM selects a deployment. When a request is labeled (e.g., think), the hook changes the model from whatever Claude Code requested to the corresponding alias, allowing seamless redirection to different models.



The diagram shows how routing labels (⚡ default, 🧠 think, 🍃 background) map to their corresponding model deployments:



```mermaid

graph LR

    subgraph ccproxy_yaml["ccproxy.yaml"]

        R1["


rules:
- name: default
- name: think
- name: background

"]

    end



    subgraph config_yaml["config.yaml"]

        subgraph aliases[" "]

            A1["


model_name: default
litellm_params:
  model: claude-sonnet-4-5-20250929

"]

            A2["


model_name: think
litellm_params:
  model: claude-opus-4-5-20251101

"]

            A3["


model_name: background
litellm_params:
  model: claude-3-5-haiku-20241022

"]

        end



        subgraph models[" "]

            M1["


model_name: claude-sonnet-4-5-20250929
litellm_params:
  model: anthropic/claude-sonnet-4-5-20250929

"]

            M2["


model_name: claude-opus-4-5-20251101
litellm_params:
  model: anthropic/claude-opus-4-5-20251101

"]

            M3["


model_name: claude-3-5-haiku-20241022
litellm_params:
  model: anthropic/claude-3-5-haiku-20241022

"]

        end

    end



    R1 ==>|"⚡ default"| A1

    R1 ==>|"🧠 think"| A2

    R1 ==>|"🍃 background"| A3



    A1 -->|"alias"| M1

    A2 -->|"alias"| M2

    A3 -->|"alias"| M3



    style R1 fill:#e6f3ff,stroke:#4a90e2,stroke-width:2px,color:#000



    style A1 fill:#fffbf0,stroke:#ffa500,stroke-width:2px,color:#000

    style A2 fill:#fff0f5,stroke:#ff1493,stroke-width:2px,color:#000

    style A3 fill:#f0fff0,stroke:#32cd32,stroke-width:2px,color:#000



    style M1 fill:#f8f9fa,stroke:#6c757d,stroke-width:1px,color:#000

    style M2 fill:#f8f9fa,stroke:#6c757d,stroke-width:1px,color:#000

    style M3 fill:#f8f9fa,stroke:#6c757d,stroke-width:1px,color:#000



    style aliases fill:#f0f8ff,stroke:#333,stroke-width:1px

    style models fill:#f5f5f5,stroke:#333,stroke-width:1px

    style ccproxy_yaml fill:#e8f4fd,stroke:#2196F3,stroke-width:2px

    style config_yaml fill:#ffffff,stroke:#333,stroke-width:2px

```



And the corresponding `config.yaml`:



```yaml

# config.yaml

model_list:

  # aliases here are used to select a deployment below

  - model_name: default

    litellm_params:

      model: claude-sonnet-4-5-20250929



  - model_name: think

    litellm_params:

      model: claude-opus-4-5-20251101



  - model_name: background

    litellm_params:

      model: claude-3-5-haiku-20241022



  # deployments

  - model_name: claude-sonnet-4-5-20250929

    litellm_params:

      model: anthropic/claude-sonnet-4-5-20250929

      api_base: https://api.anthropic.com



  - model_name: claude-opus-4-5-20251101

    litellm_params:

      model: anthropic/claude-opus-4-5-20251101

      api_base: https://api.anthropic.com



  - model_name: claude-3-5-haiku-20241022

    litellm_params:

      model: anthropic/claude-3-5-haiku-20241022

      api_base: https://api.anthropic.com



litellm_settings:

  callbacks:

    - ccproxy.handler

general_settings:

  forward_client_headers_to_llm_api: true

```



See [docs/configuration.md](docs/configuration.md) for more information on how to customize your Claude Code experience using `ccproxy`.



## Routing Rules



`ccproxy` provides several built-in rules as an homage to [claude-code-router](https://github.com/musistudio/claude-code-router):



- **MatchModelRule**: Routes based on the requested model name

- **ThinkingRule**: Routes requests containing a "thinking" field

- **TokenCountRule**: Routes requests with large token counts to high-capacity models

- **MatchToolRule**: Routes based on tool usage (e.g., WebSearch)



See [`rules.py`](src/ccproxy/rules.py) for implementing your own rules.



Custom rules (and hooks) are loaded with the same mechanism that LiteLLM uses to import the custom callbacks, that is, they are imported as by the LiteLLM python process as named module from within it's virtual environment (e.g. `import custom_rule_file.custom_rule_function`), or as a python script adjacent to `config.yaml`.



## Hooks



Hooks are functions that process requests at different stages. Configure them in `ccproxy.yaml`:



| Hook                 | Description                                                                         |

| -------------------- | ----------------------------------------------------------------------------------- |

| `rule_evaluator`     | Evaluates rules and labels requests for routing                                     |

| `model_router`       | Routes requests to appropriate model based on labels                                |

| `forward_oauth`      | Forwards OAuth tokens to providers (supports multi-provider with custom User-Agent) |

| `forward_apikey`     | Forwards `x-api-key` header to proxied requests                                     |

| `extract_session_id` | Extracts session ID from Claude Code's `user_id` for LangFuse tracking              |

| `capture_headers`    | Logs HTTP headers as LangFuse trace metadata (with sensitive value redaction)       |



Hooks can accept parameters via configuration:



```yaml

hooks:

  - hook: ccproxy.hooks.capture_headers

    params:

      - headers: ["user-agent", "x-request-id"] # Optional: filter specific headers

```



See [`hooks.py`](src/ccproxy/hooks.py) for implementing custom hooks.



## CLI Commands



`ccproxy` provides several commands for managing the proxy server:



```bash

# Install configuration files

ccproxy install [--force]



# Start LiteLLM

ccproxy start [--detach]



# Stop LiteLLM

ccproxy stop



# Check proxy server status (includes url field for tool detection)

ccproxy status         # Human-readable output

ccproxy status --json  # JSON output with url field



# View proxy server logs

ccproxy logs [-f] [-n LINES]



# Run any command with proxy environment variables

ccproxy run  [args...]

```



After installation and setup, you can run any command through the `ccproxy`:



```bash

# Run Claude Code through the proxy

ccproxy run claude --version

ccproxy run claude -p "Explain quantum computing"



# Run other tools through the proxy

ccproxy run curl http://localhost:4000/health

ccproxy run python my_script.py



```



The `ccproxy run` command sets up the following environment variables:



- `ANTHROPIC_BASE_URL` - For Anthropic SDK compatibility

- `OPENAI_API_BASE` - For OpenAI SDK compatibility

- `OPENAI_BASE_URL` - For OpenAI SDK compatibility



## Development



### Request Lifecycle



```mermaid

sequenceDiagram

    participant CC as cli app

    participant CP as litellm request → ccproxy

    participant LP as ccproxy ← litellm response

    participant API as api.anthropic.com



    Note over CC,API: Request Flow

    CC->>CP: API Request
(messages, model, tools, etc.)

    Note over CP,LP: 



    Note right of CP: ccproxy.hooks.rule_evaluator

    CP-->>CP: ↓

    Note right of CP: ccproxy.hooks.model_router

    CP-->>CP: ↓

    Note right of CP: ccproxy.hooks.forward_oauth

    CP-->>CP: ↓

    Note right of CP: 

    CP->>API: LiteLLM: Outbound Modified Provider-specific Request



    Note over CC,API: Response Flow (Streaming)

    API-->>LP: Streamed Response

    Note right of CP: First to see response
Can modify/hook into stream

    LP-->>CC: Streamed Response
(forwarded to cli app)

```



### Local Setup



When developing ccproxy locally:



```bash

cd /path/to/ccproxy



# Install in editable mode with litellm bundled

# Changes to source code are reflected immediately without reinstalling

uv tool install --editable . --with 'litellm[proxy]' --force



# Restart the proxy to pick up code changes

ccproxy stop

ccproxy start --detach



# Run tests

uv run pytest



# Linting & formatting

uv run ruff format .

uv run ruff check --fix .

```



The `--editable` flag enables live code changes without reinstallation. The handler file (`~/.ccproxy/ccproxy.py`) is automatically regenerated on every `ccproxy start`.



**Note:** Custom `ccproxy.py` files are preserved - auto-generation only overwrites files containing the `# AUTO-GENERATED` marker.



## Troubleshooting



### ImportError: Could not import handler from ccproxy



**Symptom:** LiteLLM fails to start with import errors like:



```

ImportError: Could not import handler from ccproxy

```



**Cause:** LiteLLM and ccproxy are in different isolated environments.



**Solution:** Reinstall ccproxy with litellm bundled:



```bash

# Using uv tool (from PyPI)

uv tool install claude-ccproxy --with 'litellm[proxy]' --force



# Or from GitHub (latest)

uv tool install git+https://github.com/starbased-co/ccproxy.git --with 'litellm[proxy]' --force



# Or for local development (editable mode)

cd /path/to/ccproxy

uv tool install --editable . --with 'litellm[proxy]' --force

```



### Handler Configuration Not Updating



**Symptom:** Changes to `handler` field in `ccproxy.yaml` don't take effect.



**Cause:** Handler file is only regenerated on `ccproxy start`.



**Solution:**



```bash

ccproxy stop

ccproxy start --detach

# This regenerates ~/.ccproxy/ccproxy.py

```



### Verifying Installation



Check that ccproxy is accessible to litellm:



```bash

# Find litellm's environment

which litellm



# Check if ccproxy is installed in the same environment

$(dirname $(which litellm))/python -c "import ccproxy; print(ccproxy.__file__)"

# Should print path without errors

```



## Contributing



I welcome contributions! Please see the [Contributing Guide](CONTRIBUTING.md) for details on:



- Reporting issues and asking questions

- Setting up development environment

- Code style and testing requirements

- Submitting pull requests



Since this is a new project, I especially appreciate:



- Bug reports and feedback

- Documentation improvements

- Test coverage additions

- Feature suggestions

- Any of your implementations using `ccproxy`