Open source
The security proxy for AI agents
Your agent can access every API key in plaintext — and you have no idea what it costs or where requests go. Unclaw is a forward proxy that intercepts all traffic, injects secrets without exposing them, and shows you everything. Works with OpenClaw, Claude Code, Codex, or any agent — no code changes.
Built by Deno
The problem
Secrets in plaintext
Your OpenClaw gateway token, GitHub PAT, Slack credentials — all in plaintext. Skills can access them. One prompt injection and they're exfiltrated.
Low visibility
Your agent talks to GitHub, Slack, Anthropic, and 20 other services. Hundreds of requests per hour. You can't see what it costs or what's failing.
The problem
Your agent needs real credentials to do real work
So you hand it your secret keys
One prompt injection, one buggy loop, one logged error…
Your agents could do anything
The solution
Same agents, same APIs, same code
But the secrets live outside the agent
Agents send placeholders; Unclaw injects real credentials at the edge
Agents can't leak what they can't see
Every request is logged
Secrets go only to approved destinations
How it works
Secret injection
- Placeholders swapped for real credentials at the proxy
- Inject into headers, body, or mTLS
- Secrets never reach agent memory
- Anti-exfiltration blocks reflection attacks
- Shared secrets with per-agent access control
Analytics
- Every outbound request logged in real time
- LLM costs, tokens, cache hits, and latency across all services
- Drill into full headers, body, and formatted prompts
JavaScript policy engine, AI proctoring, and human-in-the-loop approval.
What you've been missing
See everything your agents do in the Unclaw dashboard
Thousands of requests across dozens of services. Unclaw captures it all passively, with zero instrumentation.
^ Real data from one agent, one day.
How it compares
More than a gateway, more than a sandbox
AI gateways see your model calls. Sandboxes isolate your process. Unclaw does both — it sees every request and controls what credentials your agent can use.
| Secret injection | All outbound traffic | Understands LLM traffic | Handles webhooks | Analytics | |
|---|---|---|---|---|---|
| Helicone | ✕ | ✕ | ✕ | ||
| Portkey | ✕ | ✕ | ✕ | ||
| LiteLLM | ✕ | ✕ | ✕ | ||
| agentgateway | ✕ | ✕ | ✕ | ||
| Clawvisor | ✕ | ✕ | ✕ | ||
| httpjail | ✕ | ✕ | ✕ | ✕ | |
| Unclaw |
Built-in plugins
Plugins are pre-configured integrations with external services. Connect your agent(s) without writing the request-handling, auth, or secret-management code yourself.
— OR —
Write your own plugin in one TypeScript fileOpen source
The proxy handles your secrets — it must be auditable. MIT licensed. Multiple agents share secrets and endpoints, each with their own policies. Self-host or use unclaw.dev.
npm install -g unclaw