agent-kit is for the part of running AI agents that the demo skips. First-party SDKs give you a capable loop; agent-kit gives you the controls, the cost ceiling, and the record that let you put that loop in production, and hand the evidence to an auditor.
Govern What Agents Do
- >Hooks and Approval Gates: Block tools, require human approval, redact tool output, or stop a run. Fail-closed, and every decision is audited
- >Tool Allowlists: Enforced at call time, not just hidden from the prompt
- >Tool-Output Scanning: Hidden instructions, chat-template tokens, exfiltration links, and known-malicious URLs, domains, and hashes (via Nullcone) are blocked or marked untrusted before the model reads them
- >Agents as Tools: Sub-agents inherit the caller's policy, budget, and approvals, so they can't be used to route around a rule
Prove What Happened
- >Tamper-Evident Audit Chain: A hash-linked record of every model call, tool call, and policy decision, verified locally and re-verified server-side. A parent run's chain commits to each sub-agent's chain
- >Signed Evidence Bundles: Ed25519-signed exports with key rotation that anyone can verify offline with
agent-kit verify - >Retention, Legal Holds, Deletion Receipts: Holds block purges; every purged run gets a signed receipt
Control Cost and Failure
- >Cost Circuit Breaker: Per-run caps and daily/weekly/monthly fleet budgets stop an agent before its next model call. Caps fail closed on unpriced models
- >Durable Runs: Checkpoints every turn; resume after a crash, suspend for approval for days, never run a side effect twice
- >Circuit Breakers and Retry: Every state change lands in the audit chain
Operate the Fleet
- >Self-Hostable agent-kit Cloud: Fleet metrics, alerting (Slack, PagerDuty, webhook, SMTP), budgets, and the hosted audit trail on your own Postgres, via Docker Compose
- >Bring Existing Agents: Adapters for the Claude Agent SDK and the OpenAI Agents SDK, plus OTLP trace ingest for any OpenTelemetry-instrumented agent
- >Compliance Mapping: The README maps each control to SOC 2 and EU AI Act record-keeping evidence
Releases
v0.4.0 published to PyPI as agent-kit-ai on September 15, 2026 (trusted publishing); v0.5.1 is current. 365 tests across 18 numbered specs.
Tech Stack
Python SDK with Anthropic, OpenAI, Ollama, and any OpenAI-compatible provider behind one interface; MCP toolsets; FastAPI + Postgres ops backend with a worker and admin CLI.