Everything the kit carries.

Agent authoring

You say what the guide should teach; an agent drives the UI on your runner and captures each step with a validated resolver and a caption. You review the result like a pull request — curate, edit captions, approve.

• MCP tools: start_authoring · observe · act · expect · write_flow · heal_flow
• works with Claude or any MCP client

Deterministic replay

Flows re-run with zero LLM involvement. Same route, same assertions, every time — replay is free, so you can run every guide on every push.

• walkabout run --all · per-step timing, drift and screenshots
• verdicts: unchanged · changed · broken · could-not-run

Freshness & CI

Exit codes gate your builds; sweep history and alerts live in the web app. A not-compared verdict guards dry-runs from false greens.

• exit 0 clean · 1 changed · 2 broken · 3 could-not-run
• hosted status badge + freshness alerts

Rendering & publishing

Accepted runs render to git-diffable Markdown + PNGs and portable self-contained HTML with a project index. Publish to walkabout.run for stable URLs, keep repo-only, or both.

• visibility: private · unlisted · public
• custom domains · embeds · versioned guide URLs

Sessions & credentials

Session injection, storage-state import from your e2e suite, or first-party credentials staged in the Keychain and filled cross-origin-guarded. Flows hold names only — never values.

• walkabout session capture | import | refresh
• walkabout credential stage · encrypted 0600 · swept by tests

Self-hosted runners

The runner is a small package on any Mac. It connects outbound-only to pull jobs and push artifacts — no inbound ports, so internal and staging apps stay internal.

• token pairing · health & last-seen · per-project routing
• fixture/launch mode keeps personal data out of native-app guides