`.cabinet` — manifest schema

Every cabinet folder has a .cabinet file at its root. It's the only file Cabinet requires. Without it, a folder is just a folder; with it, the folder becomes a cabinet that the app can open, the registry can install, and other cabinets can nest.

schemaVersion: 1
id: my-startup
name: My Startup
description: One-person SaaS company with an AI team.
kind: root
version: 0.1.0
entry: index.md

Every field

Field	Type	Required	Default	What it does
`schemaVersion`	integer	yes	—	Always `1` right now. Lets Cabinet evolve the schema without breaking old cabinets.
`id`	string	yes	—	Slug used as the cabinet's stable identifier. `[a-z0-9-]+`.
`name`	string	yes	—	Display name. Shown in the sidebar header and in the registry.
`description`	string	no	""	One-line summary. ≤80 chars — that's the registry card limit.
`kind`	`"root"` \| `"child"`	yes	—	Root cabinets stand alone; child cabinets nest under another cabinet's folder.
`version`	semver	no	`"0.1.0"`	Useful for templates published to cabinets.sh.
`entry`	path	no	`"index.md"`	The page Cabinet opens when you click into the cabinet.
`icon`	string	no	inherits	Path or URL to a square icon used in the sidebar and registry.
`tags`	string[]	no	`[]`	Free-form tags used for registry filtering.
`license`	SPDX id	no	inherits	`MIT`, `Apache-2.0`, etc. Shown on the registry page.
`author`	object	no	—	`{name, url, email}`. Shown on the registry page.
`providers`	object	no	—	Provider routing config. See BYOAI for the shape.
`dispatch`	object	no	—	Cross-department dispatch policy. See Org chart.
`approvals`	object	no	—	Auto-approval rules for low-impact proposals.

Root vs. child

A root cabinet is the top of a tree. It has its own .cabinet, lives directly in $CABINET_DATA_DIR, and shows up in cabinetai list.

A child cabinet is a folder inside another cabinet that gets its own scope, agents, and settings. The child has its own .cabinet, and the parent treats it as a sub-cabinet (it shows up nested in the sidebar).

~/cabinets/
├── my-startup/                ← root
│   ├── .cabinet               ← kind: root
│   └── clients/
│       └── acme/
│           ├── .cabinet       ← kind: child
│           └── index.md
└── job-hunt-hq/               ← root
    └── .cabinet               ← kind: root

Provider routing example

providers:
  defaults:
    lead: claude-opus-4-7
    specialist: claude-sonnet-4-6
  fallbacks:
    claude-opus-4-7:
      - claude-sonnet-4-6
      - gpt-4.1
  budgets:
    daily:
      maxCostUsd: 25
    perTask:
      maxCostUsd: 5
  localOnly: false

Subkey	Purpose
`defaults.lead` / `defaults.specialist`	Default model per agent type.
`fallbacks`	Per-model fallback chain when the primary returns 429 or hits a budget cap.
`budgets.daily.maxCostUsd`	Hard cap per cabinet per day.
`budgets.perTask.maxCostUsd`	Hard cap per individual task.
`localOnly`	If true, refuse any non-local provider.

Dispatch policy example

dispatch:
  allow:
    - from: marketing/gtm-lead
      to: research/*
      kinds: [LAUNCH_TASK]
    - from: operations/ops-coordinator
      to: marketing/copywriter
      kinds: [LAUNCH_TASK, SCHEDULE_JOB]
  acceptFromParent:
    - from: ops-coordinator
      kinds: [LAUNCH_TASK]

See Org chart & departments for the full story.

Approvals example

approvals:
  auto:
    - kind: LAUNCH_TASK
      maxCostUsd: 0.10
      fromAgent: linkedin-operator
      toAgent: copywriter
    - kind: SCHEDULE_TASK
      maxCostUsd: 0.50
      fromAgent: ceo

Anything outside policy queues for manual approval. There's no global "trust everything" toggle.

Validation

Cabinet validates the manifest at boot. cabinetai doctor re-runs validation. A bad manifest exits with code 6 and a precise field-level error.

Read on

Persona schema — the agent file format.
Job schema — the routine YAML format.
Skill schema — the SKILL.md format.
File structure — what else lives in a cabinet folder.

.cabinet — manifest schema