Using ccproxy as an LLM API Server

ccproxy exposes an OpenAI-compatible and Anthropic-compatible API via a mitmproxy-based interceptor. Any SDK or HTTP client that supports custom base_url can use it.

Installation

System-wide (Home Manager)

Add ccproxy as a flake input and enable the Home Manager module:

nix
# flake.nix
inputs.ccproxy.url = "github:starbaser/ccproxy";

# home configuration
programs.ccproxy = {
  enable = true;
  settings = {
    # Override defaults here (port, providers, transforms, etc.)
  };
};

This installs the ccproxy binary, generates sibling ccproxy.yaml and LiteLLM-compatible config.yaml files from Nix, and creates a systemd --user service that auto-restarts when either changes.

Standalone (any Linux)

bash
# Clone and enter devShell
git clone https://github.com/starbaser/ccproxy
cd ccproxy
nix develop   # or: direnv allow

# Initialize config
ccproxy init          # copies both templates to ~/.config/ccproxy/
ccproxy init --force  # overwrites both existing files

# Edit native services and model declarations
$EDITOR ~/.config/ccproxy/ccproxy.yaml
$EDITOR ~/.config/ccproxy/config.yaml

# Start
ccproxy start

Per-project instance

Each project can run its own ccproxy with isolated config, port, and transforms via the flake's mkConfig. Use ccproxy.defaultSettings.settings (top-level, no ${system} selector needed) as the base to inherit all defaults (hooks, shaping, providers, otel).

nix
# project flake.nix
{
  inputs.ccproxy.url = "github:starbaser/ccproxy";
  inputs.nixpkgs.url = "github:NixOS/nixpkgs/nixpkgs-unstable";
  inputs.flake-utils.url = "github:numtide/flake-utils";

  outputs = { self, nixpkgs, flake-utils, ccproxy }:
    let
      defaults = ccproxy.defaultSettings.settings;
    in
    flake-utils.lib.eachDefaultSystem (system:
      let
        pkgs = nixpkgs.legacyPackages.${system};
        proxyConfig = ccproxy.lib.${system}.mkConfig {
          settings = {
            port = 4010;  # per-project: use 4010+ to avoid collisions
            inspector = {
              port = 8090;
              cert_dir = "./.ccproxy";
            };
            lightllm = {
              transforms = [
                { match_path = "/v1/messages"; action = "redirect";
                  dest_provider = "anthropic"; dest_base_url = "https://api.anthropic.com";
                  dest_path = "/v1/messages"; }
              ] ++ defaults.lightllm.transforms;
            };
          };
        };
      in {
        devShells.default = pkgs.mkShell {
          packages = with pkgs; [
            ccproxy.packages.${system}.default
            just process-compose
          ];
          shellHook = proxyConfig.shellHook;
        };
      });
}

mkConfig generates Nix-store ccproxy.yaml and config.yaml files; its shellHook symlinks both into .ccproxy/ and exports CCPROXY_CONFIG_DIR. The .envrc just needs use flake.

Add .ccproxy/ to .gitignore — the directory contains a Nix-generated symlink that is machine-specific and regenerated on nix develop:

# .gitignore
.ccproxy/

Port assignment conventions

| Port | Use | |------|-----| | 4000 | System-wide ccproxy (Home Manager, default) | | 4001 | ccproxy project's own devShell | | 4010+ | Per-project instances | | 8083 | System inspector UI (default) | | 8084 | ccproxy dev inspector | | 8090+ | Per-project inspector UI |

Running the instance

bash
# Foreground
ccproxy start

# Via process-compose (recommended for dev)
just up       # process-compose up --detached
just down     # process-compose down

# Check health
ccproxy status              # Rich panel
ccproxy status --json       # Machine-readable
ccproxy status --proxy      # Exit 0 if proxy up, 1 if down
ccproxy status --inspect    # Exit 0 if inspector up, 2 if down

process-compose.yml

Use ccproxy status --proxy as the readiness probe so dependent processes wait for the proxy to be healthy:

yaml
# process-compose.yml
version: "0.5"

processes:
  ccproxy:
    command: "ccproxy start"
    readiness_probe:
      exec:
        command: "ccproxy status --proxy"
      initial_delay_seconds: 5
      period_seconds: 30
      timeout_seconds: 10
      failure_threshold: 6
    availability:
      restart: on_failure
      backoff_seconds: 2
      max_restarts: 5

  myapp:
    command: "python -m myapp"
    depends_on:
      ccproxy:
        condition: process_healthy

Wiring SDK clients

Point any SDK at the per-project port with a sentinel key:

python
import anthropic

client = anthropic.Anthropic(
    api_key="sk-ant-oat-ccproxy-anthropic",
    base_url="http://localhost:4010",  # per-project port
)

Or via environment variables in shellHook / .envrc:

bash
export ANTHROPIC_BASE_URL="http://localhost:4010"
export ANTHROPIC_API_KEY="sk-ant-oat-ccproxy-anthropic"

Configuration

Configuration lives under $CCPROXY_CONFIG_DIR (default ~/.config/ccproxy/): ccproxy.yaml owns native services and config.yaml owns LiteLLM-compatible model declarations.

yaml
ccproxy:
  host: 127.0.0.1
  port: 4000

  providers:
    anthropic:
      auth:
        type: command
        command: "jq -r '.claudeAiOauth.accessToken' ~/.claude/.credentials.json"
      base_url: https://api.anthropic.com
      path: /v1/messages
      type: anthropic
    gemini:
      auth:
        type: command
        command: "jq -r '.access_token' ~/.gemini/oauth_creds.json"
      base_url: https://cloudcode-pa.googleapis.com
      path: "/v1internal:{action}"
      type: gemini

  hooks:
    inbound:
      - ccproxy.hooks.inject_auth
      - ccproxy.hooks.extract_session_id
    outbound:
      - ccproxy.hooks.inject_mcp_notifications
      - ccproxy.hooks.verbose_mode
      - ccproxy.hooks.shape

  shaping:
    enabled: true
    shapes_dir: ~/.config/ccproxy/shapes

  inspector:
    port: 8083
    cert_dir: ~/.config/ccproxy

  lightllm:
    transforms:
      - match_path: /v1/messages
        action: redirect
        dest_provider: anthropic
        dest_base_url: https://api.anthropic.com
        dest_path: /v1/messages

See reference/routing-and-config.md for transform rules, providers patterns, and hook parameters.

How authentication works

OAuth mode (subscription accounts -- Claude Max, Team, Enterprise):

1. Client sends sentinel key sk-ant-oat-ccproxy-{provider} as API key
2. inject_auth hook detects sentinel prefix, looks up real token from providers[name].auth
3. shape hook replays a captured {provider}.mflow shape: strips configured headers, injects content_fields from the incoming request, runs shape inner-DAG hooks (UUID regeneration, Anthropic billing-header re-signing, cache breakpoint normalization), stamps the result onto the outbound flow
4. Request reaches provider API with valid OAuth Bearer token and full identity envelope (user-agent, anthropic-beta, x-stainless-, billing header, system prompt prefix)

API key mode (direct API keys):

1. Client sends real API key via x-api-key or Authorization header
2. Key passes through to the provider unchanged

Sentinel key format

sk-ant-oat-ccproxy-{provider}

Where {provider} matches a key in providers config. Common values:

• sk-ant-oat-ccproxy-anthropic -- uses providers.anthropic.auth token
• sk-ant-oat-ccproxy-gemini -- uses providers.gemini.auth token

Default hooks

yaml
hooks:
  inbound:
    - ccproxy.hooks.inject_auth
    - ccproxy.hooks.extract_session_id
  outbound:
    - ccproxy.hooks.gemini_cli
    - ccproxy.hooks.inject_mcp_notifications
    - ccproxy.hooks.verbose_mode
    - ccproxy.hooks.shape
    - ccproxy.hooks.commitbee_compat

• inject_auth -- substitutes sentinel key with real token, sets Authorization: Bearer {token} (or the custom auth.header), clears other auth headers, and stamps ccproxy auth metadata for routing/retry
• extract_session_id -- parses metadata.user_id for MCP notification routing
• gemini_cli -- wraps Gemini sentinel-key bodies in the v1internal envelope, conditionally masquerades google-genai-sdk/* UAs, rewrites paths to cloudcode-pa.googleapis.com
• inject_mcp_notifications -- injects buffered MCP terminal events as tool_use/tool_result pairs
• verbose_mode -- strips redact-thinking-* from anthropic-beta to enable full thinking output
• shape -- replays a captured shape ({provider}.mflow) onto the outbound flow, stamping identity headers, billing header, and system prompt prefix
• commitbee_compat -- last-mile compatibility shim for the commitbee tool

AuthAddon and GeminiAddon are full mitmproxy addons (not pipeline hooks) registered after the outbound stage: AuthAddon handles 401 detection / refresh / replay; GeminiAddon handles capacity fallback + cloudcode-pa envelope unwrap.

Shape replay -- where identity comes from

ccproxy does not synthesize Claude Code identity headers in code. Anthropic-bound traffic depends on a shape: a real mitmproxy.http.HTTPFlow from the Claude CLI persisted as a .mflow file. ccproxy ships a packaged default shape for Anthropic; a user-captured shape at ~/.config/ccproxy/shapes/anthropic.mflow overrides it. The shape hook replays the shape on every outbound flow, providing user-agent, anthropic-beta, x-stainless-, the signed x-anthropic-billing-header, and the system prompt prefix.

If the shape in effect is from an outdated Claude CLI release, Anthropic will reject the request with 401/400. Capture (or refresh) a local override with:

bash
ccproxy run --inspect -- claude -p "shape capture"
ccproxy shapes save anthropic

See docs/shaping.md for the canonical reference (capture workflow, shape inner-DAG hooks, billing salt configuration, custom hooks).

Quick start

python
# Anthropic SDK (OAuth via sentinel key)
import anthropic
client = anthropic.Anthropic(
    api_key="sk-ant-oat-ccproxy-anthropic",
    base_url="http://localhost:4000",
)

# OpenAI SDK
from openai import OpenAI
client = OpenAI(
    api_key="sk-ant-oat-ccproxy-anthropic",
    base_url="http://localhost:4000",
)

SDK integration

Anthropic Python SDK

python
import anthropic

client = anthropic.Anthropic(
    api_key="sk-ant-oat-ccproxy-anthropic",
    base_url="http://localhost:4000",
)

response = client.messages.create(
    model="claude-sonnet-4-5-20250929",
    max_tokens=1024,
    messages=[{"role": "user", "content": "Hello"}],
)

No extra headers needed -- the shape hook replays the captured Anthropic shape, supplying anthropic-beta, anthropic-version, the signed billing header, and the system prompt prefix automatically.

Streaming:

python
with client.messages.stream(
    model="claude-sonnet-4-5-20250929",
    max_tokens=1024,
    messages=[{"role": "user", "content": "Hello"}],
) as stream:
    for text in stream.text_stream:
        print(text, end="")

OpenAI Python SDK

python
from openai import OpenAI

client = OpenAI(
    api_key="sk-ant-oat-ccproxy-anthropic",
    base_url="http://localhost:4000",
)

response = client.chat.completions.create(
    model="claude-sonnet-4-5-20250929",
    messages=[{"role": "user", "content": "Hello"}],
)

Requires a transform rule to rewrite from OpenAI format to the destination provider format via lightllm.

LiteLLM SDK

python
import asyncio, litellm

async def main():
    response = await litellm.acompletion(
        model="claude-sonnet-4-5-20250929",
        messages=[{"role": "user", "content": "Hello"}],
        api_base="http://127.0.0.1:4000",
        api_key="sk-ant-oat-ccproxy-anthropic",
    )
    print(response.choices[0].message.content)

asyncio.run(main())

Note: litellm.anthropic.messages bypasses proxies. Always use litellm.acompletion().

Claude Agent SDK

python
import os
os.environ["ANTHROPIC_BASE_URL"] = "http://localhost:4000"
os.environ["ANTHROPIC_API_KEY"] = "sk-ant-oat-ccproxy-anthropic"

from claude_agent_sdk import query, ClaudeAgentOptions

async for message in query(
    prompt="Your prompt here",
    options=ClaudeAgentOptions(
        allowed_tools=["Read", "Glob"],
        permission_mode="default",
        cwd=os.getcwd(),
    ),
):
    # Handle AssistantMessage, ResultMessage, etc.
    pass

Environment variables (any SDK)

bash
export ANTHROPIC_BASE_URL="http://localhost:4000"
export ANTHROPIC_API_KEY="sk-ant-oat-ccproxy-anthropic"
# OpenAI compat
export OPENAI_BASE_URL="http://localhost:4000"
export OPENAI_API_BASE="http://localhost:4000"

curl (raw HTTP)

bash
curl http://localhost:4000/v1/messages \
  -H "Content-Type: application/json" \
  -H "x-api-key: sk-ant-oat-ccproxy-anthropic" \
  -H "anthropic-version: 2023-06-01" \
  -d '{
    "model": "claude-sonnet-4-5-20250929",
    "max_tokens": 100,
    "messages": [{"role": "user", "content": "Hello"}]
  }'

Model routing

Normal model routing comes from LiteLLM-compatible model_list declarations in config.yaml. Selection order is: explicit lightllm.transforms override, exact compiled model binding, wildcard compiled binding, then sentinel-key Provider fallback. Transform overrides remain the edge-case escape hatch for host/path/model regex matching. Unmatched reverse proxy flows get a 501 error; unmatched WireGuard flows pass through unchanged.

See reference/routing-and-config.md for transform configuration patterns.

Troubleshooting

Authentication failures are the most common issue. Follow this decision tree:

Error message?
│
├─ "This credential is only authorized for use with Claude Code"
│  ▶ See: Missing or stale captured shape (system prompt prefix not stamped)
│
├─ "OAuth is not supported" / "invalid x-api-key"
│  ▶ See: Missing or stale captured shape (anthropic-beta not stamped)
│
├─ 401 Unauthorized / token errors
│  ▶ See: Token issues
│
├─ Connection refused / timeout
│  ▶ See: Connectivity
│
└─ Other / unclear
   ▶ See: General diagnostics

See reference/troubleshooting.md for the full diagnostic guide with resolution steps for each branch.

Quick diagnostic commands

bash
ccproxy status              # Verify proxy is running
ccproxy status --json       # Machine-readable status with URL
ccproxy logs -f             # Stream logs in real-time
ccproxy logs -n 50          # Last 50 lines

Known limitations (upstream flake issues)

1. Shape required for Anthropic — there is no synthetic-identity fallback. If the packaged default (or a user-captured override at ~/.config/ccproxy/shapes/anthropic.mflow) is stale for the current Claude CLI release, requests fail with 401/400. Refresh via ccproxy shapes save anthropic.
2. devConfig overwrites inspector atomically — top-level // merge on inspector drops sub-keys not re-specified. Deep merge each nested attrset explicitly: defaults.inspector // { ... }.
3. supportedSystems limited — only x86_64-linux and aarch64-linux; aarch64-darwin not supported.

Reference files

• reference/troubleshooting.md -- Full diagnostic decision tree with error-specific resolution steps
• reference/routing-and-config.md -- Model routing, config.yaml patterns, hook pipeline details