Gatekeep

When the security review asks what your agent did — show them.

Approval gates and tamper-evident audit evidence for AI-agent actions. Cards render inside Claude, ChatGPT, and VS Code; every decision lands in an append-only, hash-chained log your customer's security engineer can verify without trusting us. Self-hosted.

Book a 15-min call The eight questions ↓
▶  2-minute demo — replace this block with your video embed

Every enterprise security review asks the same eight questions.

If you sell AI agents into banks, health systems, or anyone with a vendor-risk team, you've seen these. Gatekeep exists to answer them — with evidence, not paragraphs.

1 · Authorization proofWho authorized this action — and can you prove it was that specific human?
2 · Human oversightWas a person genuinely in control? Could they have said no and stopped it?
3 · What & which ruleWhat exactly was proposed, and which policy allowed or blocked it?
4 · Segregation of dutiesCan the requester and the approver be the same identity?
5 · Log integrityCan these records be altered, truncated, or deleted after the fact?
6 · ReplayCan you rebuild any single past decision, step by step, on demand?
7 · OutcomeDid the approved action actually execute — and what happened?
8 · Data handlingWhat's stored, where does it live, and how long is it kept?

How it works

1 — Gate
Your agent calls request_approval before consequential actions. Policy decides: auto-approve within limits, or route to a human — as an interactive card inside the AI surface they already use.
2 — Evidence
Every request, decision, and outcome is appended to a hash-chained, engine-enforced append-only log — with the frozen policy and the exact context the decision saw.
3 — Prove
Export the evidence with a standalone verifier any security engineer can run. The Agent Trust Report turns your log into the answers their questionnaire is asking for.
The honest line: tamper-evident, never "tamper-proof." Edits and deletions are detectable by anyone holding an export; identity hardening (OIDC) and external anchoring are on the public roadmap. Vendors who claim more than their architecture supports are the ones your reviewer distrusts — so we don't.

Built for design partners right now

Five agent-vendor teams, shaping the evidence layer around their real security reviews. First 30 days free, then $500/month, month-to-month, self-hosted in your infra. Grab a slot →