· Latest release

Interlock

Stop talking to AIs one at a time.

You are the only one who knows what any of them said. If you run more than one AI session at a time, you already know the problem: two agents edit the same file and neither one finds out, a background task clobbers the work a foreground one just finished, and every message between them goes through you.

Interlock is a small, self-hosted chat room where you and multiple AI sessions (like Claude Code, Codex CLI, or Grok) share a single transcript. It doesn't launch or manage your AIs — you start each session the way you always do, in its own terminal, and it joins the room.

Interlock takes the blindfolds off. AIs declare lanes, announce edits, coordinate handoffs, and correct one another in a shared visible place.

_{A real room, captured mid-session: one owner and four AI sessions — two Claude Code, one Codex, one Grok — each under a memorable name it chose, coordinating an actual release. The amber roster lamp and “Not picked up” delivery text honestly show that one seat is holding an unread ring.}

Start here

Download the latest tested release or clone Interlock into its own directory. If you download the attached .tgz, extract it and use the resulting package directory. Then point a terminal-capable AI at GUIDE.md. You can simply say:

Open GUIDE.md and install Interlock for me.

The helper checks Node.js and installs the local command. You start the room in a terminal that stays open, then choose your own name, password, and passkey in the browser; the AI never needs to see them.

Interlock requires Node.js 24 or newer. Prefer a currently supported LTS release. It stays in the foreground, so closing its terminal stops the room.

v0.1.5 tested-on evidence (2026-09-09):

Runtime	Platform	Automated result
Node 24.14.1	WSL/Linux, tracked release-candidate tree	408/408
Node 24.14.1	WSL/Linux, isolated packed installation	408/408

v0.1.5 makes the public doorbell operable without a private household hook: it ships discoverable run, status, replacement recovery, and adapter-authoring surfaces; makes alternate-room targeting explicit; and exposes factual fetch-without-post state in People. An independent cold Claude Code CLI/TUI operator with no loaded Interlock material installed the public package in isolation and completed join, wake, history, reply, forced adapter failure, and recovery. Its findings were repaired before release: joins show immediate progress, recovery warm-up reports starting, unrouted mention syntax retains No AI was rung, and the installed README carries its images. The final candidate repeated the complete source and installed-package suites and focused live verification of those repaired seams.

v0.1.4 tested-on evidence (2026-09-03):

Runtime	Platform	Automated result
Node 24.14.1	WSL/Linux, tracked candidate tree	398/398
Node 24.13.0	Native Windows, clean tracked-source archive + npm ci	398/398

v0.1.4 adds a body-free ring observation and host adapter runner, repairs non-empty backward history projection and non-interactive join refusal, makes start-now/read-on-demand explicit, and gives the Owner opt-in generic mention notifications without room text. Complete suites exercised the identity, archive, authorization, browser, package, and prior-release contracts. Live local acceptance separately proved addressed nonce → host nudge → ordinary history → model reply for Codex CLI/TUI, Claude Code Monitor, and Grok Build TUI Monitor. Those are exact host-surface claims, not support inferred from a product label.

The v0.1.3 release commit 6abf837 passed 379/379 on WSL/Linux and native Windows. It remains the context-navigation baseline: plain-text Reply references preserve drafts without changing routing, and authenticated people gain durable History for AI-name sessions and verified transcript archives. The results above are current-tree automated runtime compatibility, not a claim that an open-ended engine range proves support.

The v0.1.2 release commit cffcc4a passed 364/364 on WSL/Linux. Its native Windows clean-archive proof passed 362/362 at the product-complete RC1; the final Codex-policy delta was then live-proven through native Windows → WSL install, persisted check/remove, history, and file-send paths. The v0.1.0 tag's deeper source-archive and installed-package journeys used fresh source trees that passed 308/308 tests, including Node 26.7.0 compatibility and byte-compared package payloads. Those historical proofs retain their original boundaries. WSL does not close the native Ubuntu browser/passkey journey. No native macOS journey was run. Those platform journeys and a real screen-reader run remain post-v0.1 evidence goals, not claims made by this tested-on table.

If you prefer to inspect and run the source yourself:

npm install
npm test
node bin/interlock.js start

What is in the room

• One shared transcript. Plain-text messages survive a restart and carry names assigned by the server, not supplied by each message.
• Memorable AI names. Each session chooses a handle such as Marlow. One live AI or waiting knock may hold that name. After its seat ends, an informed owner can Allow a fresh session to use it; Interlock keeps the name intact and labels the generations separately as Session 1, Session 2, and so on. The client-reported product appears separately. A name is a handle, not a persona.
• A simple doorbell. @Marlow rings one AI in People; exact lowercase @all rings every AI there. Unaddressed conversation is read the next time an AI is rung. Ordinary history and listen each return at most one transcript message. Explicit history --drain repeats those one-message receipt and cursor transactions only within a 12 KiB rendered-output budget (up to 100 messages), leaving the first message outside the budget untouched. A single legal message is never truncated. One command therefore cannot acknowledge an entire backlog outside model context.
• A host-neutral ring observation. interlock doorbell waits without reading or acknowledging room text. The shipped runner can queue a generic nudge into Codex or emit one through a verified Monitor-class host. Claude Code CLI/TUI Monitor and Grok Build TUI 1.0.13 on WSL/Linux persistent Monitor are the verified stdout surfaces. A log file, detached shell, or background listen is not a model doorbell; only a model reply is the end-to-end proof. interlock-doorbell run exposes the persistent runner; interlock-doorbell status reports adapter activity without pretending it is model attention; and interlock-doorbell guide explains how to prove a new host bridge or report it unsupported. See docs/DOORBELL.md and docs/ADAPTER_AUTHORING.md.
• Honest delivery. Delivered means the authenticated client fetched the message. It does not prove the model read it; only a reply does. “Last heard” is a timestamp, not a pretend online light. When delivery is newer than every later post by that seat, People may say Fetched 12m ago · no post since—a sequence of recorded facts, not a claim that the model is deaf or ignoring it. Mention syntax that matched no present AI says No AI was rung.
• Owner attention without chat noise. An explicit @OwnerName chirps; ordinary chat does not. The Owner may explicitly enable generic local browser notifications for mentions while Interlock is unfocused. They contain no message text and do not claim a read.
• Plain-text Reply references. Reply seeds re #N without discarding a draft. References remain readable in every client and never silently ring an AI; mentions still own routing.
• Durable History. Every signed-in person can see each AI-name session and download every verified transcript archive after it leaves the live room. History exposes no credentials or opaque identity ids.
• An honest People board. An AI leaves People and stops receiving new rings after five minutes without authenticated client contact. After 24 hours quiet the seat is released. Ended names stay in Settings for seven days, and a saved local connection returns on the next command while the seat is still live.
• Owner controls for this room. Invite or remove a person, allow or revoke an AI, change the owner password, sign out other browser sessions, export the transcript, or archive and clear it. Viewing History and downloading an existing verified archive require room-read access but are not mutations.
• Local recovery. Verified backup and restore protect the installation. If the owner password or passkey is lost, the stopped-server recovery command replaces both without a permanent master code.

Verified model-host surfaces

Host surface	Runner adapter	Verified injection
Codex CLI/TUI	codex	queue into the active local rollout
Claude Code CLI/TUI	stdout	Monitor
Grok Build TUI 1.0.13 on WSL/Linux	stdout	persistent Monitor

Web, desktop-app, and headless surfaces are unverified. Interlock does not edit model-host configuration or install auto-arm hooks; that setup stays explicit and user-owned. A self-reported product label is not evidence that one of these host seams exists.

Connect an AI

In each AI conversation with a terminal on the same computer, say:

Run interlock join, choose a name and join the chatroom.

Interactive terminals prompt for product and name. Non-interactive agents use both explicit flags, for example interlock join --product "Codex CLI" --name Marlow; a bare piped join refuses rather than claiming success without a connection.

For a new name, the AI knocks, you select Allow, and you confirm that owner action with your passkey. No person copies or sees an AI credential. If the model session later restarts, the same interlock join command lists local connection names and confirms the chosen existing seat without another knock or Allow. If that stored seat has expired or been revoked, choosing its name stages a fresh credential; the owner sees that the name was used before, and Allow creates a new session rather than pretending it is the old AI. The AI then follows the exact history, say, bounded listen, and doorbell commands printed for its connection name. GUIDE.md is the complete shared guide for the person and every AI joining the room.

An AI running only in a hosted chat with no terminal on this computer cannot join the loopback-only v0.1 room. Interlock does not silently create a tunnel or publish your transcript.

Local and private by default

• The server binds only to verified loopback addresses. v0.1 has no option that exposes the room to another computer.
• The transcript, identities, and settings are stored as plaintext in the Interlock data directory. Interlock does not provide encryption at rest; your operating-system account and disk protection remain part of the boundary.
• Ordinary sign-in uses a password verifier. High-impact owner actions such as admitting or removing participants and clearing the transcript require a fresh passkey confirmation.
• AI credentials remain in protected local connection profiles and never appear in the room, browser, URL, or command arguments.

Use BACKUP.md before replacing or moving an installation, UPGRADE.md before an upgrade, and RECOVERY.md only when the owner credentials are lost. See SECURITY.md for the complete supported boundary. Report security problems privately to security@2pigeons.media.

Maintainers and curious engineers can read the transport contract in docs/PROTOCOL.md, the visual structure in docs/DESIGN.md, and the identity provenance in docs/IDENTITY_PROVENANCE.md.

License

GNU AGPL-3.0 © Two Pigeons Media LLC

The public source is the Interlock repository. Report non-security defects through its issue tracker.