The receipt.

A stranger’s first 0dai run, recorded end to end — real account, real repo, real receipt. No demo magic, no staged output. Every number below is read from the run’s own artifact, which you can download and replay.

Recorded 2026-07-04 · live mode · agent claude· the account was a throwaway, the repo was not ours.

The evidence

Project context was injected into the run.

Before the agent changed a line, 7project files — 3,669bytes — were read in as context. Not a template: the actual manifests and docs the CLI found in the repo.

  • ai/manifest/project.yaml
  • ai/manifest/discovery.json
  • ai/manifest/commands.yaml
  • ai/manifest/current_state.json
  • ai/manifest/current_task.json
  • ai/docs/roadmap.md
  • ai/docs/decisions.md
36sto first productive result (from init marker to run start)
7context files
3,669bytes of context
$0.30est. cost

Receipt status: partial · score 33 · agent claude. That label is not a hedge — here is exactly what “partial” means.

What actually ran

Eight steps, install to receipt, with the real time each one took. Only one was driven by a human; the run step fanned out into the agent tasks the receipt counts.

  1. Install the CLI40.2s
  2. Sign in(human browser login — a real person entered the device code)117.1s
  3. Activate the free tier2.5s
  4. Detect the stack and write project context19.5s
  5. Run the health check26.6s
  6. Run the task183.1s
  7. Capture the receipt0.0s
  8. Scan every command for operator-only calls0.0s

Why the labels are honest

“Partial” is the true label, not a soft “pass”

Of 3 agent tasks, 1 changed a file (pass), 2 exited clean without changing anything (noop), and 0failed. A noop is never counted as a pass. Pass, noop, and failed are separate states — the score summarizes the run, it does not rubber-stamp it.

No operator-only calls, checked by scanning every command

All 7 commands the run issued were scanned for operator-only or admin-only calls. None were found. A first run uses only the self-service commands any new signup can run — no internal endpoint, no admin token.

A fresh home directory, so nothing could ride on our session

The run used a fresh home directory and a throwaway account, so it could not inherit any existing sign-in on the machine. Authentication started from zero, exactly as it would on a stranger's laptop.

Replay it yourself

You do not have to trust the receipt — reproduce it. With a supported agent CLI installed and signed in, run the same sequence against your own small repo:

npx -y @0dai-dev/cli auth login
npx -y @0dai-dev/cli activate free
npx -y @0dai-dev/cli init
npx -y @0dai-dev/cli doctor
npx -y @0dai-dev/cli run "Add a NOTES.md file describing this project" --now --yes

The full replay instructions — including the throwaway-account and fresh-home setup that keeps the run honest — live in the public repository: github.com/0dai-dev/0dai.

Rather inspect the raw recording? Download the exact artifacts this page is built from: the scored receipt and the full run artifact (every step, every command, the operator-only scan result).

See what ships today on status and the security posture on the trust center.