For builders

Show what your agent did, to someone who doesn't trust you.

Agents act for people and companies. Logs can't be shown to a stranger. Screenshots can be edited. "Trust our dashboard" doesn't travel. When something goes wrong, or the other side disputes it, nobody can show what was asked, what was approved and what was done.

What is different

A book you keep, rules you can read.

  • Each action is sealed as it happens, into a book you keep.
  • Terms Packs you can read say what must be shown.
  • A judge is labelled as a judge. Its call is never mixed with the facts.
  • Anyone can verify the book offline, without us.
  • When both sides keep a book, their halves can be compared.

Record → Rules → Check → Result → View. Build at one step; the rest still works.

The detail of each step is in the open standard: agentactioncapsule.org ↗

How little it takes

Three levels. Start at any one.

Level 1

Receipts in one step

Install Action State Receipts. Your agent's deals and actions get receipts, sealed on your machine. The install uses capsulectl, a release candidate build, and runs on Linux x86_64.

Quickstart →

Level 2

Add a Terms Pack

Check your agent's records against the EU AI Act Obligations pack. The pack says what must be shown, and every judged answer is marked as judged.

Quickstart ↗

See a record judged, end to end: Demos →

Check your work

Verify without us.

Records verify against public keys, offline with scitt-cose ↗ or your own code. Checkpoints can go to the free public witness ↗, which sees hashes, never content.