A new project called openclaw-claude-bridge (published on GitHub about 6 hours ago by shinglokto) solves a problem that's been quietly frustrating OpenClaw power users: getting extended thinking, persistent session memory, and granular cost tracking for Claude without juggling raw API keys or worrying about rate limits directly.
The bridge is an OpenAI-compatible HTTP proxy that sits between OpenClaw and Claude Code CLI. Instead of OpenClaw calling the Anthropic API directly, it calls the bridge, which handles the Claude Code CLI session management, extended thinking, and exposes a React dashboard with real-time per-agent cost tracking.
Why This Architecture Is Interesting
After Anthropic blocked OAuth access for third-party tools in early 2026, OpenClaw users were left with two options: direct API keys (with their own rate limits and cost tracking challenges) or the official Claude Code CLI (which has better tooling but was designed for coding workflows). The bridge merges the best of both.
What you get that you don't with direct API keys:
- Persistent sessions: The bridge maintains Claude Code CLI sessions across OpenClaw restarts. Conversations can resume rather than starting fresh each time.
- Extended thinking exposed properly: Claude Code CLI's extended thinking mode is surfaced as the
thinkingparameter in the OpenAI-compatible API, so OpenClaw's/thinkcommand works correctly with it. - Per-agent cost tracking: The React dashboard shows cost broken down by which OpenClaw agent drove the spend — useful if you're running multiple agents or billing clients.
- Auto-resume: If a session drops, the bridge auto-resumes from the last checkpoint rather than losing context.
Setup (Early — Expect Rough Edges)
Heads up: This project is new — published today. It's working software but early-stage. Treat it as a power-user tool, not a production-ready turnkey solution. Read the README and watch the GitHub issues before deploying it in anything critical.
Prerequisites: Claude Code CLI installed and authenticated (claude --version should work), Node.js v18+, and an existing OpenClaw setup.
# Clone and install git clone https://github.com/shinglokto/openclaw-claude-bridge cd openclaw-claude-bridge npm install # Start the bridge npm start # Bridge runs on http://localhost:3099 by default
Then configure OpenClaw to use the bridge as an OpenAI-compatible provider:
# In ~/.openclaw/openclaw.json
"models": {
"providers": {
"claude-bridge": {
"baseUrl": "http://127.0.0.1:3099/v1",
"apiKey": "bridge",
"api": "openai-completions",
"models": [
{
"id": "claude-opus-4-6",
"name": "Claude Opus 4.6 (via Bridge)",
"contextWindow": 200000
}
]
}
}
}
Set it as your default model:
/model claude-bridge/claude-opus-4-6
The React dashboard runs on http://localhost:3099/dashboard and shows per-session and per-agent cost in real time.
Pros and Cons vs. Direct API Keys
✅ Pros
- ✓Persistent sessions across restarts
- ✓Extended thinking works correctly
- ✓Per-agent cost dashboard
- ✓Auto-resume on session drop
- ✓One place to manage Claude auth
⚠️ Cons
- ✗Extra process to keep running
- ✗Very new — limited community support
- ✗Claude Code CLI must be installed and authed
- ✗Adds latency (small but measurable)
- ✗Not an official Anthropic integration
Who Should Try This
This setup makes most sense for:
- Multi-agent setups where you want to see which agent is costing what
- Long-running sessions where conversation continuity across gateway restarts matters
- Power users who already have Claude Code CLI set up and want to centralize their Anthropic auth in one place
It's probably overkill for single-agent setups with simple workloads. Direct API keys are simpler and fine for most use cases.
Worth watching: This is exactly the kind of community tooling that tends to become standard practice once it matures. Bookmark the repo and check back in a few weeks — early projects like this often get significantly more polished quickly when they catch community attention.
Alternatives If You Don't Want the Extra Complexity
If the bridge feels like too much overhead, the standard direct API key setup works fine. The main things you lose are session persistence across restarts (which you can partially work around with good workspace files) and the per-agent cost dashboard (which you can approximate with Anthropic's usage console).
For most single-agent setups, direct API keys + good SOUL.md + a solid memory.md are all you need. The bridge is a power-user tool, not a requirement.