Documentation

A sealed file for every agent run

EPI turns an AI agent run into a portable .epi evidence file: steps, hashes, Ed25519 seal, optional policy analysis, optional SCITT / timestamp. You verify offline — no dashboard required to trust the seal.

The loop

Record — record() or wrap OpenAI / LangChain / LiteLLM while the agent runs.
Seal — SHA-256 of members + Ed25519 over the canonical manifest → one .epi.
Verify — epi verify or browser private check (file not uploaded).
View — epi view opens a forensic timeline in any browser.

Three different questions

01 / SEAL

Tampered?

Integrity + signature of this file. Valid seal ≠ known who.

02 / IDENTITY

Who sealed?

KNOWN (org pin) / LOCAL / NOT_PINNED. Unpinned → WARN, not claim-ready. epi keys trust · --policy strict

03 / POLICY

Rules broken?

Optional epi_policy.json + deterministic analyzer. Separate from seal. epi analyze

What a .epi is

Envelope-v2 polyglot: valid HTML + ZIP payload. Contains manifest.json, hash-chained steps.jsonl, environment, optional analysis, embedded viewer. Change one byte and the signature fails.

Verify modes

Claims / insurers — epi verify file.epi --policy strict (org-pinned sealer required).
Offline CLI (dev) — epi verify file.epi · unpinned identity is WARN · UNVERIFIED IDENTITY.
Private browser — file stays in the tab · /verify/
Share — epi share publishes a hosted link to the sealed file; hosted deep-dive reports are part of the control plane (design partners).

Depth (GitHub)

Try on home Open verify tool →