A flight recorder for AI agents

Your AI agents act.
Can you prove what they did?

AI is starting to do real things — approve refunds, move money, take actions for your customers. When one gets it wrong, “our logs say so” isn’t proof: your logs can be edited. Swarrm keeps an unfakeable record of what your AI actually did — so you can prove it to a customer, an auditor, or a court. And anyone can check it’s real, trusting no one.

a record, checked by anyone — offline, no account
$ swarrm verify refund-2041.json
   this record was made by the agent it claims  (signature ok)
   it hasn’t been changed or reordered since   (sealed & chained)
   the action: refund_bot approved $240 refund, 14:02 UTC
RESULT: VERIFIED

$ # now someone quietly edits one number…
$ swarrm verify refund-2041-edited.json
   the record no longer matches its seal
RESULT: NOT VERIFIED

Why “check our logs” doesn’t cut it

Ordinary logging turns into “just trust us” on the one day it actually matters.

Your logs can be edited

Whoever owns the database owns the story. In a dispute, “trust our logs” is the first thing the other side throws out.

The AI vendor can’t vouch for itself

A platform grading its own AI’s homework isn’t proof. Neutral, independent evidence is something the big AI vendors structurally can’t sell you.

Proof has to survive the wait

Disputes, audits, and regulators show up months later. Evidence has to convince someone who wasn’t there — or it isn’t evidence.

Three steps. One line of code to start.

Think of it as a black box for AI: it quietly records, and it never gets in the way.

01 — CONNECT

Point your AI at Swarrm

One line of code, one command, or the monitoring you already run. It never blocks or slows your AI — and if Swarrm goes down, your AI keeps working. That’s a hard rule, not a setting.

02 — IT RECORDS

Every action, sealed

Each thing your AI does becomes a sealed record. We keep a fingerprint, never your prompts, tools, or customer data. Records are chained so nothing can be quietly changed or deleted — not even by us.

03 — PROVE IT

Anyone can check it

Export one file. Anyone can confirm it’s genuine — right in their browser, no account, no cooperation needed from you or from us. Ever.

Don’t take our word for it

Drop a real record into your browser and check it yourself — nothing gets uploaded. Then change one character and watch it fail. That’s the whole point: you never have to trust us.

Verify a real one →

The honest version

This only works if our own promises are exact. So, plainly:

Start free, on your own machine

No account, no cloud bill, no sales call. Install it, run the demo, and see a verified record in about two minutes.

Get started →

Straight answers

What is this, in one sentence?
A black box for AI agents: it keeps an unfakeable record of what your AI actually did, so you can prove it to anyone later — and they can check it’s real without trusting you or us.
Will it slow down or break my AI?
No. Recording happens alongside your AI, never in the way. If Swarrm is unreachable, your agents keep running and records buffer safely until it’s back — a gap even shows up honestly in the record. This is tested, not promised.
Do you see my data?
No. We store fingerprints (one-way hashes), not your prompts, tool calls, or customer data. In production the recording runs inside your own network, so your data never leaves it.
Who can verify a record?
Anyone. The verifier is open-source and runs in a plain browser — no account, no API access, no help needed from you or us. There are two independent versions that have to agree, so you’re not trusting a single black box.
Is this blockchain?
No token, no chain of ours to buy into. We optionally stamp records using a public timeline so their timing can be proven to someone who wasn’t watching — that’s it. The record verifies fine without it.
What does it cost?
Running it yourself is free. Verifying is free forever. Paid, hands-on partnerships are for teams that want us embedded in their product — reach out.