Roadmap

Cahut is built in batches, each one a working slice from the AI to the reader. Their order is set, their dates are not: the milestones of the repository say where each batch stands.

The batches

  1. Batch 0

    Foundations and decisions

    The first architecture records, the Go skeleton, the continuous integration and the release build.

    Done

  2. Batch 1

    Publish over MCP, serve over HTTPS

    An AI publishes a Window with a personal token. Cahut serves it at its own address, in a sandbox, and the call goes to the Logbook.

    Done

  3. Batch 2

    People, Rooms and the console

    People sign in; administrators of each Cahut and of the server; team Rooms with their rights; the web console and its API.

    Done

  4. Batch 3

    Connect the AIs: OAuth

    Each Cahut is the authorization server of its MCP endpoint. On a consent screen, a person grants an AI some Rooms and a scope, and sees the connected AIs in the console.

    Done

  5. Batch 4

    Complete Windows: versions and access

    Updating a Window keeps the previous version, and one step rolls back. Each Window opens to its Room only, to everyone signed in to its Cahut, to anyone with its Key, or to the public.

    Done

  6. Batch 5

    Knock first

    An AI asks before anything it cannot undo. The request waits in the console, with a notification and an expiry, until a person approves or declines it.

    Done

  7. Batch 6

    House style and a skill for Claude

    The organization's colors, fonts, logo and templates, given to every AI, and a skill that tells Claude how to publish on Cahut.

    Done

  8. Batch 7

    Installation, first start and updates

    One command on a fresh AlmaLinux or Ubuntu server, a guided first start in the browser, and updates from the releases, with a rollback.

    Done

  9. Batch 8

    A second AI client and acceptance

    A second AI client tested from end to end, a security review, and the success criteria checked on a fresh server.

    Done

  10. Batch 9

    E-mail from each Cahut

    Each Cahut sends e-mail through its own SMTP relay: links to choose a password, invitations to read a Window, and notices of Knocks.

    Done

The milestones on GitHub

What works today

Batches 0 to 9 are done, and the first version met the success criteria on a fresh server. In a Cahut, today:

  • One command that installs a Linux server, a first start in the browser, and daily updates from signed releases, with a rollback.
  • One service on a Linux server, holding one or more Cahuts, each at its own domain with its certificate, or at the IPv4 address of the server until it has one.
  • Collaborative Cahuts, or individual ones where each person works alone, and limits on people, Rooms, Windows and disk, set by the administrators of the server.
  • People who sign in, administrators of the Cahut and of the server, and team Rooms with owners and members. The page of each Room lists its Windows, or shows the one its owners choose.
  • A web console, in English and French: Knocks, Rooms, Windows, the Logbook, people, connected AIs, personal tokens, e-mail and the server.
  • AI clients that connect over OAuth, such as claude.ai, Claude Code, ChatGPT and Le Chat, to the Rooms and the scope a person grants, or with a personal token.
  • Windows published over MCP: documents, decks, dashboards and static sites, served in a sandbox that holds them to what their AI declares they load and call.
  • Two versions of each Window with a rollback, a trash, and an access: Room only, internal, with a Key or public, until a date.
  • Knock first: what an AI cannot undo, or what changes what a person may do, waits for its person, who approves or declines it in the console.
  • E-mail through the SMTP relay of each Cahut: links to choose a password, invitations to read a Window, and notices of Knocks, of the account and of limits.
  • A house style for each Cahut: its colors, fonts, logo and guidance, templates for documents, decks and dashboards, and a deck engine, which AIs read before they publish.
  • A plugin for Claude Code, and a skill for the Claude apps, that teach Claude to make Windows in the house style.
  • The Logbook of every call, every change and every sign-in.

No backups yet: keep snapshots of your server, and Cahut is not ready for work you cannot lose.

The documentation keeps the detail, area by area:Where Cahut stands.

Ready for launch when

CriterionThreshold
Install on a fresh VPSUnder 10 minutes, one command
First publication by an AIUnder 5 minutes, without touching the server
Sharing by a non-technical personA deck shared by link in a few minutes
Supported AIsThe same Cahut works with at least two different AI clients
GovernanceEvery AI action is in the Logbook, and is either reversible or was approved

After the first version

Then, towards the full scope of the launch, in about this order:

  1. Comments, search, single sign-on, and backups and restore.
  2. Apps in unprivileged containers: runtimes, a Dockerfile mode, databases, secrets, logs and resource limits.
  3. Export to PDF and PowerPoint, SSH access limited to the same operations, and other AI clients.