Know every agent. Prove every action. Stop any of them. Nothing inside your agents changes.
Take the ten-question passport check
Start in watch-only mode — install, see your agents, block nothing. Switch enforcement on when you're ready. Sign in with Google or GitHub, or your work email. Already have an account? Sign in ↗
Agents borrow human credentials and share API keys. Impersonating one is trivial — and nobody can prove which agent did what.
Access is all-or-nothing. An agent allowed to read can often also write, delete, and spend. Delegation hands over far too much.
No tamper-evident record exists — nothing that would satisfy your security team, let alone a regulator asking questions.
Agent sessions burn orders of magnitude more tokens than chat, and each vendor caps spend inside its own account. Nobody ties spend to a job, an outcome and a person across all of them.
And six more your auditor will ask. → The ten-question passport check
OpenAI evaluation agents ran inside Hugging Face production systems for days — nobody could say which caller to cut. Two disclosures, one lesson: agents without identities cannot be stopped one at a time.
Hugging Face disclosure, Jul 2026 ↗A coding agent deleted a production database during an explicit code freeze, then reported the change as successful. It had one credential for everything, so it could touch everything.
The Register, Jul 2025 ↗Moffatt v. Air Canada: the airline was held liable for its chatbot's invented refund policy. "A separate legal entity" was not accepted — your agent's word is your word.
2024 BCCRT 149 ↗// none of these was a model failure — each one is an identity and authorisation failure
Every agent gets a verifiable identity anchored in your own domain — who built it, who runs it, and what it's trusted to do. Bound to the agent's own key, so a stolen copy is useless. Your systems check it in milliseconds; you revoke it in one click.
A doorman in front of the systems you already run. Your rules on every call: least privilege, human sign-off where it matters. Delegation only narrows — suspend a parent and every helper it sponsored stops within minutes.
A tamper-evident record of every action — who, on whose authority, under which rule. Each entry fingerprints the one before it, so the past can't be rewritten quietly. Built to become a regulator-ready evidence pack — in design with EU auditors now.
What every agent costs you, before the invoice arrives — one number across every vendor and account. The budget belongs to the job, not the vendor. Walk into a renewal knowing what your estate actually consumes.
▶ Watch the 2-minute tour · 1:56 · light and dark cuts
Run the installer on a machine you own. It puts a local gateway — a door your agents call through — in front of the tools they already reach. Point a program at it and its calls flow through KnowMyAgents. Nothing inside your agents changes.
Shadow Agents. Every unidentified caller KMA sees — the crew inside your app, the script nobody remembers — surfaced the moment it makes a call, before it has a passport. Including the ones nobody authorised.
The places your agents reach appear on their own. Approve or refuse each one — no allow-lists to maintain.
Give each agent a passport — a verifiable identity it presents at the gateway on every call. Shadow becomes identified. No passport, no access — if you say so.
"Only agents with a passport may call." Per-door allow or refuse. A daily spend cap. Write it in plain words and KMA drafts it for you — Draft rules with AI.
"Never trade defense-sector stocks." Judged for intent by an AI judge on every call it covers — so it catches what a hardcoded list can't.
Rules decide access deterministically. Charters add a judged conduct check on top and never widen access.
Watch first: every decision recorded as "would allow" or "would refuse", nothing blocked. Then flip to enforce — allowed calls pass, breaches are refused, and every decision is written down with its reason.
Proof. Every decision recorded, tamper-evident, with the reason — including why the judge refused. One click to an evidence pack for your security review or a regulator.
Your data stays yours. The full audit record — every request and reply detail — lives on your own machine, in the Local Viewer. The hosted console sees only metadata: who called, which door, allowed or refused, why, what it cost. The content never leaves your network.
Enforcement never depends on the console. If it goes dark, your gateways keep working.
THE CONSOLE IS A WINDOW · THE GATEWAY IS THE DOORSign-up, one-command install, doors, Shadow Agents, passports, rules and charters, observe → enforce, proof, and the Local Viewer — the eight steps above, on the real screens.
Prefer to read? The eight steps above are the tour, in words. Want the case behind it? The argument ↓
A simulation of the checks the doorman runs on every real call: is there a passport, is it genuine, is it revoked, what do the rules say.
Pick an agent and send its request to the doorman. Watch it work.
This is what the first afternoon looks like — no sales call needed.
Platform, DevEx and security teams. Your coding-agent rollout is stuck in security review, or your first surprise token bill just landed. Within a week you can answer who, what, and how much.
Agent platforms, ISVs and AI product companies. Your enterprise deals stall on one security questionnaire: what can your agents do, and how would you prove it? Give every agent you ship a verifiable passport, and hand your buyer a record instead of an assurance — without changing your product.
Point one coding-agent session at the doorman and watch what appears. Most teams find agents they never authorised inside the first hour.
One command. Point a program at it. Nothing else to deploy.
One click per agent in the console, anchored in your own domain. Shadow becomes identified.
Watch-only: every decision recorded as would allow or would refuse. Nothing blocked. Zero risk.
Flip the rules live when you're ready. Breaches are refused, and the security review finally has an answer.
irm https://knowmyagents.com/install.ps1 | iex
curl -fsSL https://knowmyagents.com/install.sh | sh
It asks where to install, then asks for one thing — a registration token from your console.
Re-run it any time to update — your secrets, registration, audit log and passports are kept.
Read the full install guide — prerequisites, updating, backup, troubleshooting ↗// one command to install · one click per passport · no SDK, nothing added to your agents
The same identity rails the EU digital wallet runs on — applied to agents. Anchored in DNS and cryptography, the trust roots the internet already accepts. Your agent traffic and your log never leave your own infrastructure. EU residency for the hosted control plane is on the roadmap — ask us where it runs today.
The EU AI Act names what this layer produces. Article 12 requires automatic recording of events; Article 14 requires oversight that can interrupt a system — which is what revoking a passport does. Articles and dates, kept current → the compliance mapping.
Every access system we built assumes a person is behind the keyboard. Agents break that assumption quietly — they act continuously, spawn helpers, and hold credentials issued to someone else. None of the incidents above was a model failure. Each one was an identity and authorisation failure. The fix is not smarter agents. It is papers.
Roles for owners, admins, policy authors, auditors and viewers. Invite colleagues. Single sign-on (OIDC).
Start free, on your own. Install the doorman with one command, and see every agent and every place it reaches — in watch-only mode, blocking nothing. Switch enforcement on when you're ready. Sign in with Google or GitHub, or your work email.
Prefer a person? Request a demo.
Prefer a walkthrough first? Write to hello@primethoughts.com.