Agentic Tools

Two rules shape it. The console never touches a project's source — the one thing it writes there is the shared task list, by design. And the terminals it drives are hosted outside the console's own process, so reloading the thing you are looking at does not kill the sessions you are watching. It ships with no build step and no framework at all — a console you cannot start because its bundler broke is not a console. A change to it is only finished once the running process has been proven to have loaded it, which is a different claim from the tests passing.

DMS — a content-addressed library that finished output is filed into, with a watcher catching what nobody filed: full-text search, version rollback, and tagged provenance on everything filed at source. Over 23,600 artifacts across 35,700 versions are stored and searchable today, in daily use on the projects that are actively building.

Capture is deliberately doubled up. Anything produced can be filed explicitly, with a title, tags and a record of what it was made for and why — and a watcher picks up whatever nobody filed, because a library that depends on remembering to use it is a library with holes. Re-filing the same thing with new bytes makes a new version rather than a replacement, and every earlier version stays retrievable byte for byte. Identical bytes are recognised and stored once. Nothing gets pruned: the store is past twenty-nine gigabytes and the standing decision is to warn rather than delete, because disk is cheaper than a deliverable nobody can find again.

WhatsApp Bridge — links Matt's own WhatsApp numbers as a tool an AI can read, search and send from, several numbers at once, each with its own rate limits. A send needs an explicit recipient, explicit text and an explicit go — never inferred, never automatic.

The interesting part of that bridge is what it refuses. Which number a message goes out from is never guessed either — the sending number is as much a decision as the recipient. Sends on one number are put through a queue rather than run in parallel, because the rate check reads the same ledger the send writes, and two callers passing the same check at the same time is exactly how a burst gets through a limit. A repeat of the same send key returns the original result instead of sending twice, across every linked number; a send that was refused leaves its key usable, so a corrected retry is not blocked by the mistake. And what arrives in a message is treated as information, never as an instruction, whoever it appears to be from.

Social Studio — a cross-brand social engine where the AI drafts posts and schedules them, and the operator approves before anything reaches a real audience. Once approved, the publish itself runs unattended at the scheduled time.

Approval there is granted at the item, not at the moment. The operator clears a batch of drafts and the engine then publishes them on schedule without asking again — because an approval step that fires at midnight is an approval step that gets rubber-stamped. Nothing unapproved ever publishes, replies to comments are drafted for a one-tap send, and following and liking are left entirely to the operator: no interface exists for them, and driving the site through a browser to fake one risks the account the whole thing depends on. Even the publish order is fixed, because the image has to be reachable before the post that points at it goes out.

Jarvis Whisper — a Windows dictation overlay for push-to-talk and hands-free voice typing, built on a speech-to-text pipeline that was benchmarked against real captures and tuned on the results rather than picked on reputation.

Its model choices are settled by replay, not by release notes. A candidate is run over real captures already on disk and compared against what the live system produced; the last one tested came back measurably faster and less accurate, and was not adopted. A speech detector runs locally and decides when anyone is actually talking, and the system will only load an approved detector by name and path, and a local transcriber stays loaded as the fallback for when the fast path is unavailable. The overlay and the runtime are separate processes, so the visible half can be restarted without dropping what is being said.