How Cabinet works, in precise terms.

For developers and anyone vetting Cabinet. Each section starts with the short version, then says what the product does today. If that's not you, the home page says the same in plainer words.

Rules

A deterministic policy engine, not the model, decides every action. Personas only propose, and agents can't change the policy. Low-risk actions run within daily caps. Changes to budgets, models, tools, or policy wait for Approve and Reject buttons, and pending approvals expire after 7 days.

Every kind of action has a ceiling set in policy: it runs on its own, runs after a rule check, needs an approval, or is denied. A workspace's own settings can only narrow those ceilings, never widen them.

Posting to a team channel, writing to the ledger, filing a GitHub issue, and handing work to the engineer run on their own, each within a daily limit. Changing settings or the rules, making a hire, and cancelling engineering work need an approval.

Admins and owners can approve. Some approvals are an owner's alone: a change to the budget, to a persona, to the company settings, or to the merge limits or engineering autonomy, a hire, and cancelling work. A reason typed with Reject goes to the team member that asked, if its meeting or task is still running.

Meetings and decisions

Meetings run blind positions, anonymized revision (up to two rounds by default), and a synthesis by the chair. The skeptic joins when agreement runs high and doesn't vote. Each decision record holds the options, an owner, a confidence, the dissent, a falsifier, and a review date, and a retro grades it worked, mixed, or failed.

The chief of staff chairs. It restates the question, lists the decision criteria and what is out of scope, says whether web research is needed, and picks participants by what each owns: by default three to five, never itself and never the skeptic.

Each participant writes a position on its own: a recommendation, a rationale, risks, a confidence, and what would change its mind. A model then groups the recommendations and measures agreement. By default the skeptic joins when more than 80% back the same course, argues the strongest case against it, and does not vote. A workspace can set one to four revision rounds, and a team that is split asks the owner which way to go.

The decision is a numbered Markdown record in the MADR 4.0 format: context, options, the decision, consequences, confidence, falsifier, dissent, actions with an owner each, and a review date. Later a retro grades it worked, mixed, or failed.

Files

Decisions, minutes, retros, playbooks, and persona configuration are Markdown files in a ledger, which can sync both ways with a repository in your GitHub account.

The ledger lives in the database first, so Cabinet works with no repository, and reads never wait on GitHub. With the GitHub App installed on your organization, it syncs both ways with a repository, which Cabinet can create. If both sides changed the same file, GitHub's version wins and the earlier one stays in history.

Changes that arrive from GitHub to a team member's instructions or the company notes wait for an owner to review and apply them.

Engineering

Nightshift works through its GitHub App in the repositories you choose and opens pull requests, hosted by Cabinet or, on Pro, self-hosted. Work items carry a brief and acceptance criteria. An independent review checks each change, and autonomy tiers set by your plan decide what may merge without a person.

Nothing is built until you hire Nightshift and connect GitHub. The merge level is set by the plan: on Trial and Owner no change merges without a person; on Team, low-risk changes merge after review; on Pro you choose which changes may merge. Merge size limits (files and lines) and the checks a repository requires are workspace settings, and only an owner can change the limits.

The reviewer uses a different model provider from the builder wherever another provider is available. After a merge, a deploy watch follows the repository's deploy workflow.

Hosted work runs on Cabinet. On the Pro plan, work can run on your own servers instead.

Models

A router sends each task to Claude, OpenAI models, Gemini, or Grok, with fallback between them. Included usage runs on Gemini by default, and on Claude when Gemini is unavailable. Your own Anthropic, OpenAI, Google, or xAI API keys work on every plan at the same price, and use none of the included usage.

Each kind of task has its own ordered list of models, and the router moves to the next one when a model is unavailable.

When the included usage is spent, the morning update, the weekly meeting, and decision follow-ups keep running on a small monthly reserve, while chats, research, and hires wait until you add a balance or the month resets. Calls made with your own keys never use it.

Controls

One monthly budget cap enforced in code. Per-workspace roles (owner, admin, member, viewer), an audit trail, and provider keys stored sealed per workspace. Text from Slack, Google Chat, and web pages is treated as untrusted data, not instructions.

Four roles, ascending: a viewer reads, a member also chats, an admin also approves, changes settings, connects tools, and manages members, and an owner can do everything. The budget cap warns at 80%.

The audit trail records settings changes, switches, and approvals with who made them; admins and owners can search it.

Provider keys and the tokens of connected tools are encrypted with AES-256-GCM under a key derived for each workspace, with the workspace bound to the ciphertext, so a stored credential copied to another workspace does not open.

Connections

Slack, Google Chat, and the web for the conversation; the tools listed above for the work. Each hire works through its own connection, with only the access it needs, under rules enforced in code. Google tools ask for the scopes a tool needs at the moment you connect it.

Connected today: Slack, Square, Google Sheets, Stripe, Shopify, Google Analytics, Google Ads, Search Console, HubSpot, QuickBooks, GitHub, X. In beta: Microsoft Teams, Outlook, Excel, OneDrive, SharePoint, Xero, Faire, Twilio (they work today; the provider's review isn't finished). Owners and admins make the connections.

GitHub is read for the team (changes, issues, pull requests, and workflow runs) and written to only by a Nightshift hire, which opens pull requests in the repositories you choose. Which Google scopes each tool asks for is listed in the privacy policy.