Agent Receipt: independent proof of agent actions

A verifiable independent record of what an AI coding agent actually did

Problem

A developer pointed out that an AI agent's transcript is not ground truth: the agent may claim it edited files or ran tests when it didn't. He built an open-source project, Rashomon, that independently records shell commands, exit codes, tool calls and test results, then compares them against the agent's own account. This kind of agent audit is becoming a real anxiety for teams using tools like Claude Code.

Solution

A personal agent execution recorder: hooks into common agents (e.g. Claude Code), logs the commands actually run, files touched and tests executed, and emits a signable JSON receipt; highlights discrepancies between an agent's self-report and reality; lets users share receipts with teammates or clients in one click.

Users

Indie developers and tech leads heavily using AI coding agents in CI or locally, plus contractors who need to hand over compliance evidence.

MVP

CLI wrapper that records shell and file changes from common agents; generates human-readable execution receipts; flags obvious mismatches with the agent's narrative; stores records locally as JSON; exports receipts as Markdown.

Revenue

Open-source core plus a paid desktop tier with a team dashboard, receipt signing and compliance exports, billed monthly or per seat.

Why now

Agent-written code is now normal, but the trust problem is only just being discussed; existing tools focus on prompt management or observability, not on independently reconciling what an agent said versus did, leaving a small but painful gap.