Live browser sandbox · real proofs, simulated side effects

Proof-gated actions for AI agents

PermitRail sits between an agent and a risky tool call. It checks policy, asks for approval of the exact request, signs a short-lived proof, lets the action run once, and writes a receipt you can verify later.

Open the sandbox Read threat model

Ed25519 proofs input-bound approval single-use execution signed receipts

App integration npm install @permitrail/core @permitrail/mcp-gateway @permitrail/provider-webhook

MCP npx @permitrail/mcp

Connect your own agent · MCP client config

{
  "mcpServers": {
    "permitrail": { "command": "npx", "args": ["-y", "@permitrail/mcp"] }
  }
}

Try the API

Edit a tool call, run it, and watch the real package respond. Approve to sign a proof, execute it once, then replay it and watch it get refused.

1 Request
2 Policy
3 Approval
4 Receipt

permitrail · sandbox starting

Request exact action JSON

Example requests

Start with a scenario or paste any agent tool call. The same policy/proof flow works for your own tools too.

call permitrail_authorize_tool_call

Edit this exact request ready to edit

Change the tool, recipient, amount, purpose, or input. PermitRail binds approval to this exact JSON, not a generic permission.

Real Ed25519 signing in your browser. Nothing is sent, charged, or deleted.

Response idle

// run a call to see the real response from the package

Equivalent code

// the call you run shows up here

Call log0 calls

Every call is logged here with its status and latency.