ccproxy unlocks the full potential of your Claude MAX subscription by enabling Claude Code to seamlessly use unlimited Claude models alongside other LLM providers like OpenAI, Gemini, and Perplexity.
It works by intercepting Claude Code's requests through a LiteLLM Proxy Server, 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.
New ✨: Use your subscription without Claude Code! The Anthropic SDK and LiteLLM SDK examples in examples/ allow you to use your logged in claude.ai account for arbitrary API requests:
# Streaming with litellm.acompletion()
response = await litellm.acompletion(
messages=[{"role": "user", "content": "Count from 1 to 5."}],
model="claude-haiku-4-5-20251001",
max_tokens=200,
stream=True,
api_base="http://127.0.0.1:4000",
api_key="sk-proxy-dummy", # key is not real, `ccproxy` handles real auth
)
⚠️ Note: While core functionality is complete, real-world testing and community input are welcomed. Please open an issue to share your experience, report bugs, or suggest improvements, or even better, submit a PR!
Important: ccproxy must be installed with LiteLLM in the same environment so that LiteLLM can import the ccproxy handler.
# 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:
ccproxycommand (for managing the proxy)litellmbundled in the same environment (so it can import ccproxy's handler)
# 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.
ccproxy --help
# Should show ccproxy commands
which litellm
# Should point to litellm in ccproxy's environmentRun the automated setup:
# 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.
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.
ccproxy:
debug: true
# Optional: Shell command to load oauth token on startup (for litellm/anthropic sdk)
credentials: "jq -r '.claudeAiOauth.accessToken' ~/.claude/.credentials.json"
hooks:
- ccproxy.hooks.rule_evaluator # evaluates rules against request ─┬─ (optional, needed for
- ccproxy.hooks.model_router # routes to appropriate model ─┘ rules & routing)
- ccproxy.hooks.forward_oauth # required for claude code's oauth token
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: trueWhen ccproxy receives a request from Claude Code, the rule_evaluator hook labels the request with the first matching rule:
MatchModelRule: A request withmodel: claude-3-5-haiku-20241022is labeled:backgroundThinkingRule: A request withthinking: {enabled: true}is labeled:think
If a request doesn't match any rule, it receives the default label.
LiteLLM's proxy configuration file is where your model deployments are defined. The model_router hook takes advantage of LiteLLM's model alias feature 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:
graph LR
subgraph ccproxy_yaml["<code>ccproxy.yaml</code>"]
R1["<div style='text-align:left'><code>rules:</code><br/><code>- name: default</code><br/><code>- name: think</code><br/><code>- name: background</code></div>"]
end
subgraph config_yaml["<code>config.yaml</code>"]
subgraph aliases[" "]
A1["<div style='text-align:left'><code>model_name: default</code><br/><code>litellm_params:</code><br/><code> model: claude-sonnet-4-5-20250929</code></div>"]
A2["<div style='text-align:left'><code>model_name: think</code><br/><code>litellm_params:</code><br/><code> model: claude-opus-4-5-20251101</code></div>"]
A3["<div style='text-align:left'><code>model_name: background</code><br/><code>litellm_params:</code><br/><code> model: claude-3-5-haiku-20241022</code></div>"]
end
subgraph models[" "]
M1["<div style='text-align:left'><code>model_name: claude-sonnet-4-5-20250929</code><br/><code>litellm_params:</code><br/><code> model: anthropic/claude-sonnet-4-5-20250929</code></div>"]
M2["<div style='text-align:left'><code>model_name: claude-opus-4-5-20251101</code><br/><code>litellm_params:</code><br/><code> model: anthropic/claude-opus-4-5-20251101</code></div>"]
M3["<div style='text-align:left'><code>model_name: claude-3-5-haiku-20241022</code><br/><code>litellm_params:</code><br/><code> model: anthropic/claude-3-5-haiku-20241022</code></div>"]
end
end
R1 ==>|"⚡ <code>default</code>"| A1
R1 ==>|"🧠 <code>think</code>"| A2
R1 ==>|"🍃 <code>background</code>"| A3
A1 -->|"<code>alias</code>"| M1
A2 -->|"<code>alias</code>"| M2
A3 -->|"<code>alias</code>"| 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:
# 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: trueSee docs/configuration.md for more information on how to customize your Claude Code experience using ccproxy.
ccproxy provides several built-in rules as an homage to 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 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.
ccproxy provides several commands for managing the proxy server:
# Install configuration files
ccproxy install [--force]
# Start LiteLLM
ccproxy start [--detach]
# Stop LiteLLM
ccproxy stop
# Check that the proxy server is working
ccproxy status
# View proxy server logs
ccproxy logs [-f] [-n LINES]
# Run any command with proxy environment variables
ccproxy run <command> [args...]
After installation and setup, you can run any command through the ccproxy:
# 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 compatibilityOPENAI_API_BASE- For OpenAI SDK compatibilityOPENAI_BASE_URL- For OpenAI SDK compatibility
When developing ccproxy locally:
cd /path/to/ccproxy
# Install in development mode with litellm bundled
uv tool install --from . claude-ccproxy --with 'litellm[proxy]' --force
# After making changes, reinstall
uv tool install --from . claude-ccproxy \
--with 'litellm[proxy]' \
--force \
--reinstall-package claude-ccproxy
# Restart the proxy to regenerate handler file
ccproxy stop
ccproxy start --detach
# Run tests
uv run pytestThe handler file (~/.ccproxy/ccproxy.py) is automatically regenerated on every ccproxy start.
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:
# 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
cd /path/to/ccproxy
uv tool install --from . claude-ccproxy --with 'litellm[proxy]' --forceSymptom: Changes to handler field in ccproxy.yaml don't take effect.
Cause: Handler file is only regenerated on ccproxy start.
Solution:
ccproxy stop
ccproxy start --detach
# This regenerates ~/.ccproxy/ccproxy.pyCheck that ccproxy is accessible to litellm:
# 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 errorsI welcome contributions! Please see the Contributing Guide 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