Runtime authorization for AI agents.

Decide which agent actions are allowed to execute before they reach production systems.

   AI Agent
      │
      │  tool call
      ▼
   ┌──────────┐
   │ NullRun  │
   │   Gate   │
   └────┬─────┘
        │
        ├── ALLOW ────────────► Tool executes
        ├── REQUIRE APPROVAL ─► Human decides
        └── BLOCK ─────────────► Tool does not execute

All SDK releases →

• Tool policies — block sensitive tools before execution.
• Spend limits — hard and soft budgets, chain-aware overdraft, rate caps.
• Human approval — action-bound grants via SHA-256 payload digest.
• Workflow control — pause / kill from the dashboard over WebSocket.
• Audit trail — hash-chained audit_events for every gate decision.

NullRun evaluates structured action requests — allow, require_approval, block. It does not inspect prompts or arbitrary semantic content; tool-block policies match tool names only, and approval rules predicate over typed, explicitly forwarded tool parameters. Cost enforcement relies on SDK-reported usage — a malicious SDK that controls its own cost reports is not protected by the gate.

• nullrun-sdk-python — Python SDK (Apache-2.0, pip install nullrun).
• nullrun-docs — documentation source.
• nullrun-examples — runnable agent integrations.
• nullrun — gateway and dashboard. Private; access on request.

• Bug or feature? Open an issue in the relevant repo.
• Security? See SECURITY.md — please don't file public issues.
• Email: support@nullrun.io.