claude-proxy-rs

Unified API proxy for Claude Pro/Max subscription.

Use your existing Claude Pro/Max subscription with AI coding assistants and tools that support either OpenAI-compatible or Anthropic native APIs, including Cline, Roo Code, Kilo Code, and more.

Admin UI

Overview	API Keys	Models

Features

• Dual API support: OpenAI-compatible (/v1/chat/completions) and Anthropic native (/v1/messages)
• OAuth authentication with Claude Pro/Max subscription
• Admin UI (Vue 3 SPA) for managing OAuth, API keys, models, and usage
• Streaming support with keep-alive pings (prevents timeouts during extended thinking)
• Tool/function calling, image inputs (base64)
• Extended thinking mode (configurable via model suffix or native API parameters)
• Automatic prompt caching (auto-injects cache breakpoints for tools, system, and conversation history)
• Token counting (/v1/messages/count_tokens)
• Per-key cost-based rate limiting (5-hour/weekly/total limits in USD, synced with subscription windows)
• Per-key model access control (allow all or whitelist specific models)
• Per-model usage tracking with cost calculation (input/output/cache pricing)
• Usage history — time-series charts for cost and tokens, breakdowns by model and API key (24h/7d/30d)
• User-facing usage dashboard at /admin/usage — no admin auth required, authenticate with your sk-proxy-* key
• Dynamic model management (add/remove models, configure per-token pricing)
• Key enable/disable toggle
• Configurable cloaking mode (always/never/auto)
• Single binary deployment (admin UI embedded via memory-serve)

Installation

Requirements

• Rust >= 1.94.0 (edition 2024)
• Bun >= 1.3.0
• Vite+ (vp CLI) — unified frontend toolchain
• just command runner

Build & run

# Create .env with required admin credentials
cp .env.example .env
# Edit .env with your credentials

just build   # builds admin UI + release binary
just run     # or: cargo run

Open http://127.0.0.1:4096/admin, log in, connect your Claude account via OAuth, and generate API keys (sk-proxy-*).

Configuration

Environment variables are loaded from .env or the environment.

Admin sessions use HttpOnly cookies with a 30-day sliding expiration. Basic Auth is also accepted.

Request captures

Set CLAUDE_PROXY_CAPTURE_DIR=/path/to/captures to write one directory per inference request. Captures include redacted client headers, inbound JSON, prepared Anthropic JSON, upstream response headers, and either upstream_body.txt or raw upstream_stream.sse chunks.

Capture files may contain prompts, tool results, code, and model outputs. API keys, authorization headers, and cookies are redacted from headers, but the capture directory should still be treated as sensitive.

Data storage

All data (OAuth credentials, API keys, usage) is stored in a Turso embedded database:

• Linux: ~/.local/share/claude-proxy/proxy.db
• macOS: ~/Library/Application Support/claude-proxy/proxy.db
• Windows: %APPDATA%\claude-proxy\proxy.db

Usage

OpenAI-Compatible API

from openai import OpenAI

client = OpenAI(base_url="http://127.0.0.1:4096/v1", api_key="sk-proxy-...")

response = client.chat.completions.create(
    model="claude-sonnet-4-6",
    messages=[{"role": "user", "content": "Hello!"}]
)
print(response.choices[0].message.content)

Anthropic Native API

from anthropic import Anthropic

client = Anthropic(base_url="http://127.0.0.1:4096", api_key="sk-proxy-...")

response = client.messages.create(
    model="claude-sonnet-4-6",
    max_tokens=1024,
    messages=[{"role": "user", "content": "Hello!"}]
)
print(response.content[0].text)

Both APIs accept x-api-key: sk-proxy-... or Authorization: Bearer sk-proxy-....

IDE Extensions

Cline / Roo Code / Kilo Code (Recommended)

Use the native Anthropic provider:

Extended thinking works via the extension's built-in controls — no model suffixes needed.

Custom Tool Names

Claude subscription OAuth is sensitive to the tool names sent to the native Anthropic API. Claude Code uses a fixed set of mcp_* tool names; third-party tools often send their own names such as shell, patch, ask_followup_question, or update_todo_list.

In cloaking mode, the proxy best-effort normalizes known third-party tool names to Claude Code-compatible names before forwarding the request upstream. Tool descriptions and input schemas are kept intact, and response tool calls are mapped back to the client's original names before returning to the tool.

Unknown or colliding tool names fail locally with 400 Bad Request instead of being forwarded upstream. This is intentional: new tool vocabularies should be captured, reviewed, and added explicitly.

Current tested support:

Some aliases are semantically approximate because Anthropic only accepts the Claude Code-compatible name set. The model still sees the original tool description and schema, which are the primary signals for how to call the tool.

Alternative: OpenAI Compatible

Extended thinking suffixes (OpenAI mode only):

Opus 4.6 — adaptive thinking with effort parameter:

Older models (Sonnet 4.5, Opus 4.5, etc.) — manual thinking with budget_tokens:

Available Models

• claude-opus-4-6, claude-sonnet-4-6 (latest, adaptive thinking)
• claude-opus-4-5, claude-sonnet-4-5, claude-haiku-4-5
• claude-opus-4-1, claude-opus-4-0, claude-sonnet-4-0

API Endpoints

OpenAI-Compatible

• POST /v1/chat/completions — streaming supported
• GET /v1/models

Response extensions (ignored by standard clients):

Request extensions:

Anthropic Native

• POST /v1/messages — streaming supported
• POST /v1/messages/count_tokens
• GET /v1/models

Admin

• GET /admin — Admin UI
• GET /admin/usage — User-facing usage dashboard (Bearer key auth)

Health

• GET /health

Development

just run                  # Start backend (cargo run)

For frontend hot reload, run in a second terminal:

cd admin-ui && vp install && vp dev
# Open http://localhost:5173/admin/

Regenerating the API client

The TypeScript client in admin-ui/src/client/ is auto-generated from the OpenAPI spec. After changing backend routes or types, regenerate it with:

just openapi   # dumps spec via --openapi flag (no running server needed), then regenerates client

All recipes

just check    # fmt + clippy + tests + frontend checks
just build    # build UI + release binary
just fmt      # format Rust code
just lint     # clippy + frontend lint
just test     # Rust unit tests
just openapi  # regenerate TypeScript client
just deploy   # build + deploy to server
just logs     # tail server logs
just status   # server systemd status
just restart  # restart server service
just test-api # integration tests against running proxy