openclaw-claude-bridge

繁體中文 | English

An OpenAI-compatible HTTP proxy that lets OpenClaw agents use Claude through Claude Code CLI — with tool calling, session memory, and extended thinking.

Why This Exists

OpenClaw speaks the OpenAI API format. Claude Code CLI speaks its own format. This bridge sits in between and translates — so your OC agents can talk to Claude without either side needing to change.

  OpenClaw agent                  Bridge                     Claude CLI
  (Discord/Telegram)
        │                           │                              │
        │  POST /v1/chat/           │                              │
        │  completions              │  Translate messages           │
        ├──────────────────────────▶│  Inject tool protocol        │
        │                           │  Map thinking level          │
        │                           │                              │
        │                           │  stdin ──▶ claude --print    │
        │                           │  stdout ◀── stream-json      │
        │                           │                              │
        │  SSE stream               │  Parse response:             │
        ◀──────────────────────────│  ├─ text → clean output      │
        │                           │  └─ <tool_call> → OpenAI fmt │
        │                           │                              │
        │  tool_calls?              │                              │
        │  Yes → OC runs tools      │                              │
        │  No  → answer to user     │                              │

Key principle: The bridge never executes tools. It only translates. OpenClaw owns the tool loop.

Quick Start

# 1. Clone and install (also builds the dashboard automatically)
git clone https://github.com/shinglokto/openclaw-claude-bridge.git
cd openclaw-claude-bridge
npm install

# 2. Configure
cp .env.example .env
# Edit .env — review settings, change DASHBOARD_PASS

# 3. Make sure Claude CLI is installed and logged in
claude --version
claude auth status

# 4. Start
npm start

# 5. Health check
curl http://localhost:3456/health

How It Works

Request Flow

1. OpenClaw sends a standard OpenAI chat completion request (messages, tools, model)
2. Bridge translates the messages into Claude CLI's text format and injects tool-calling instructions into the system prompt
3. Claude CLI processes the request and streams back a response
4. Bridge parses the response — if Claude wants to call a tool, it converts the <tool_call> XML into OpenAI's tool_calls format; otherwise it returns clean text
5. OpenClaw either executes the requested tools and sends another request, or delivers the final answer to the user

Session Memory

Each agent in each conversation gets its own persistent Claude CLI session. This means Claude remembers previous messages without the bridge needing to resend the full history every time.

Discord #general — two agents sharing the same channel:

  researcher (first message)  → new session created (session-aaa)
  helper-bot (first message)  → new session created (session-bbb)
  researcher (second message) → resumes session-aaa (only sends new messages)
  user runs /new in OC        → old session purged, fresh one created

The routing key is channel + agent name, so agents in the same channel never interfere with each other.

Session state survives restarts — mappings and request history are saved to state.json and restored on startup. Stale sessions (where the CLI session file no longer exists) are automatically pruned.

For more details on the three-tier session lookup and edge cases, see docs/architecture.md.

Tool Calling

The bridge reads the tools array from OpenClaw's request and dynamically generates tool-calling instructions that get injected into Claude's system prompt. Claude outputs <tool_call> XML blocks, which the bridge converts into standard OpenAI tool_calls.

This means any new tools added in OpenClaw are automatically available to Claude — no bridge changes needed.

Claude's native tools (Bash, Read, Write, etc.) are disabled via --tools "" so it can only call tools through OpenClaw's gateway.

Extended Thinking

The bridge supports Claude's extended thinking via the reasoning_effort parameter:

When reasoning_effort is not provided, thinking is disabled entirely (MAX_THINKING_TOKENS=0).

Dashboard

The bridge includes a React dashboard accessible at http://<server-ip>:3458/.

Header bar — live metrics across the top: online/offline status, uptime, total requests, active requests, total cost, session count + disk size, error count, available tools, and a dark/light theme toggle.

Agent sidebar — lists all agents sorted by recency. Each shows an activity indicator (green if active in last 5 min, amber if 30 min, gray if idle), session count, request count, and cost. Selecting an agent filters all panels. On mobile, collapses to a horizontal pill bar.

Live activity feed — real-time event stream with emoji-coded messages: 🧠 thinking, 🔧 tool calls, 🔄 resume, ♻️ context refresh, ✅ done, ❌ error. Shows relative timestamps and agent/channel labels.

Context cards — per-session context window usage with progress bars. Color-coded: green (<40%), amber (40–65%), red (>65%). Each card shows session ID, agent, token counts, and cost.

Request table — 13-column table showing every request: time, channel, session (color-coded), resume method (emoji badges: 🔧 Tools, 💬 Chat, 🆕 New, ♻️ Refresh, etc.), prompt size, model, thinking level, input/output tokens, cost, cache hit rate, duration, and status. Rows expand to show activity logs and errors. Supports channel and resume method filtering, and pagination.

Session cleanup — one-click button to delete CLI sessions older than 24 hours.

Password protection: Set DASHBOARD_PASS in your environment to enable HTTP Basic Auth (user: admin). If not set, the dashboard is open.

For detailed architecture of the dashboard, see docs/architecture.md.

Configuration

Environment Variables

Ports

OpenClaw Setup

Add this provider to your OpenClaw config (~/.openclaw/openclaw.json):

{
  "models": {
    "providers": {
      "claude-bridge": {
        "baseUrl": "http://localhost:3456/v1",
        "apiKey": "not-needed",
        "api": "openai-completions",
        "models": [
          {
            "id": "claude-opus-latest",
            "name": "Claude Opus",
            "contextWindow": 1000000,
            "maxTokens": 128000,
            "reasoning": true
          },
          {
            "id": "claude-sonnet-latest",
            "name": "Claude Sonnet",
            "contextWindow": 1000000,
            "maxTokens": 64000,
            "reasoning": true
          }
        ]
      }
    }
  }
}

Then assign the model to your agent. The apiKey can be any non-empty string — the bridge doesn't check it.

Service Setup (Auto-Start on Boot)

macOS (launchd)

# Recommended: auto-detects paths, reads .env, generates plist
./service/install-mac.sh

Manual management:

# Status
launchctl list | grep openclaw-claude-bridge

# Restart (reload plist config)
launchctl bootout gui/$(id -u)/com.openclaw.claude-bridge
launchctl bootstrap gui/$(id -u) ~/Library/LaunchAgents/com.openclaw.claude-bridge.plist

# Logs
tail -f ~/openclaw-claude-bridge/bridge.log

Linux (systemd)

cp service/openclaw-claude-bridge.service ~/.config/systemd/user/
# Edit the file if your project path differs from ~/openclaw-claude-bridge
systemctl --user daemon-reload
systemctl --user enable --now openclaw-claude-bridge
loginctl enable-linger $USER  # start at boot without login

systemctl --user status openclaw-claude-bridge
journalctl --user -u openclaw-claude-bridge -n 50
systemctl --user restart openclaw-claude-bridge

API Reference