Skip to content

Latest commit

 

History

History

demo

README.md

The hero demo

acp-local-demo.gif is a real recording of the real thing — the live install.sh --local fetched from agenticcontrolplane.com, the real Claude Code and Codex CLIs, and denies coming from the actual hook path in a sandbox HOME. Nothing is mocked or edited. For a control-layer product, a fabricated demo would be disqualifying; if a take goes wrong, re-record it, never doctor it.

demo.cast is the asciinema source for the GIF — keep them in sync (re-render with record.sh render whenever the cast changes).

Re-recording

demo/record.sh prep          # fresh sandbox HOME + tiny demo repo (local bare origin)
demo/record.sh login-claude  # one-time: `claude setup-token` flow — macOS keeps /login
                             # OAuth in the Keychain, which the sandbox can't read, so
                             # the rig uses a long-lived token stored 600 inside .home/
demo/record.sh login-codex   # one-time; then run /hooks inside codex and TRUST the ACP hook
demo/record.sh record        # automated take (or: manual — you drive, it records)
demo/record.sh render        # demo.cast → acp-local-demo.gif (+ .mp4 for the site hero)

The sandbox isolates ~/.acp and hook config only — the agent CLIs use your normal accounts. Codex's /hooks trust step is its security model for hooks; the recording shows governance after that one-time review, which is exactly the state every real install lands in.

Episodes

record.sh <cmd> [episode] (default force-push). Each lives in episodes/<slug>.sh and defines the fixture, policy, confirm-gate, and beats.

Episode What it shows Status
force-push Hardline floor deny shipped (the hero)
replit Policy deny on a scoped infra tool shipped
claude-install Install how-to — Claude Code end to end shipped
codex-install Install how-to — Codex end to end ready to shoot

claude-install — one command, and what it writes

Terminal recording: the installer runs in local mode, then ls of the .acp directory shows decide.mjs, govern.mjs and policy.json, and grep shows the hook registered in Claude Code's settings.json. Claude Code then runs the test suite — allowed and logged — and is asked to force-push main; the human approves and the safety floor blocks it anyway. The audit log ends showing two allows and one deny.

Deliberately not another telling of the hero demo. The hero is the "what ACP does" story; this is the how-to beside it — what the one command actually wrote to your machine, how you confirm it's live, and what a governed session looks like from there. Piping a script to bash deserves scrutiny, and showing every file it touched answers that better than a paragraph can.

The closing receipt shows two allows and one deny on purpose. A log full of only denials would misrepresent what governance spends its time doing.

demo/record.sh prep claude-install
demo/record.sh record claude-install    # no human step needed
demo/record.sh render claude-install

codex-install — one command, three dialogs, first deny

Written and validated, not yet shot. Codex's install is the fiddliest of the supported harnesses: demo/shoot-codex-install.sh walks the prereqs, and the episode handles all three startup dialogs a fresh HOME hits — including the update prompt whose default is "Update now" (npm install -g @openai/codex mid-recording). It needs a human for ~20 seconds to trust the hooks, which the rig waits for and never writes itself.

Await rules, learned the expensive way

Every one of these cost a take. They apply to any episode:

  1. Never match the command you just typed. await "Ctrl|codex|>" matched the shell echoing $ codex, so the rig typed its prompt into a startup dialog.
  2. Never match text that's still on screen from an earlier beat. A block pattern of force-push to main matched the installer's own banner explaining what the floor blocks — the beat "passed" in 9 seconds without the agent running anything.
  3. Never match LLM prose. The model rephrases every run: [Tt]ests? pass missed "Yes, it passes"; a question-detector built on "Do you want me to" missed "want me to proceed with". Match deterministic tokens, or detect that the turn ended (Claude Code shows esc to interrupt while working).
  4. Scrollback answers history; the visible pane answers state. Most awaits grep capture-pane -S -60, which includes scrollback — right for "did this ever appear", wrong for "what's on screen now".
  5. Ctrl-C does not quit Claude Code, it interrupts. /exit quits, Ctrl-D is the fallback. force-push.sh and replit.sh still end with C-c; C-c and should be ported before either is re-recorded.

Prerequisites: brew install tmux asciinema agg gifsicle, ffmpeg for the mp4, and the local-mode hook fix (claude-code-acp-plugin#8) merged — without it the fetched canonical hook has no local path and the take will show nothing.

Storyline (~60s)

  1. curl -sf https://agenticcontrolplane.com/install.sh | bash -s -- --local — one command, no signup
  2. claude → asked to squash + force-push main → blocked by the safety floor
  3. same session → curl to an external API → ACP asks, declined on camera
  4. same session → run the tests → allowed, silently logged
  5. codex → the same force-push → the same floor blocks it (cross-vendor, one policy)
  6. tail -n 6 ~/.acp/audit.jsonl — the receipt: every call, its decision, its reason

The GIF embeds in this README, the org profile, and the marketing-site hero. GitHub can't autoplay video, so the GIF is the canonical embed; the mp4 is for the site (smaller, sharper).