One portal across every Clavenar surface.
Clavenar is a control plane for AI agents that inspects every tool call, runs Rego policy, parks risky ones for human approval, and hash‑chains the decision into an audit ledger. These docs cover the install path, the architecture, the wire contracts, and a cookbook of integrations.
Start from the job you need done.
Install the TS or Python wrapper, wrap your provider client, and read the first audit row.
Evaluating architectureTrace Agent -> Proxy -> Brain -> Policy -> HIL -> Ledger before any side effect fires.
Operating ClavenarRun Slack HIL, SAML, Postgres, Helm, observe rollout, and SIEM egress runbooks.
Auditing evidenceVerify a hash chain, replay one correlation ID, and map rows to compliance proof.
Run clavenar-lite or the full stack.
Use TS or Python SDK around Anthropic/OpenAI.
Agent emits a normal tool-use block.
Allow, deny, or pending review.
Correlation ID joins the chain.
Quickstart
From zero to a verdict in 5 minutes. Boot clavenar‑lite locally or on Fly.io, wrap your Anthropic / OpenAI client with the SDK, watch the ledger fill up.
Concepts
The four decision layers (Data Plane / Semantic / Governance / Forensic), the three security signals, the HIL flow, observe vs. enforce, and how the hash chain proves to an auditor what happened.
API reference
Wire‑level contracts for every endpoint the proxy, brain, policy engine, HIL, and ledger expose. JSON‑RPC envelope, headers, status codes, error shapes. Cites TECH_SPEC.md as the source of truth.
Recipe cookbook
Working examples per integration target — LangChain, Vercel AI SDK, LlamaIndex, Mastra, Anthropic Computer Use, OpenAI Realtime — plus operator runbooks for SAML, Postgres, observe rollout, Splunk egress.
Source of truth. Every per‑service wire contract is pinned in clavenar-specs/TECH_SPEC.md. These docs paraphrase + link — if a sentence here disagrees with the spec, the spec wins.