Run your work like an AI agency, from a single Notion board.
Learn how agency-os shipped - and why Haiku subagents cut your token spend 5-10x on mechanical work.
agency-os turns Notion into the dashboard of your own AI agency. You talk through ideas with the agent. It pushes back, clarifies scope, breaks the work into tasks and subtasks with the right hierarchy and dependencies, and writes it all into Notion. You approve. Agents then pick up the work and ship it - jobs, submissions, drafts, reports - with result links back on the row.
It's for founders, makers, solo operators, and teams who want their AI tools to actually ship things, not just talk about them.
- The agent plans for you. Talk through an idea in plain English. The agent asks the right questions, carves it into tasks and subtasks, sets dependencies, and writes the plan onto the board. You stop being the project manager.
- One place for everything. Ideas, tasks, decisions, and finished work all live in one board. No more chasing TODOs across Slack, docs, and twelve open tabs.
- Agents do the boring parts. Flip a task to
Exec=Agentand an AI handles it end-to-end: form fills, post drafts, directory submissions, log-and-close work. - Right model for the job. Mechanical work runs on fast, cheap models. Heavy thinking goes to bigger ones. You don't pay flagship rates for clerical tasks.
- Dependencies just work. Tasks that block other tasks run in order. The queue won't fire something whose prereq isn't done.
- You stay in control. Nothing dispatches autonomously. Every run is operator-gated; the board is honest about what's queued and why.
- Suggest an idea. It lands in the Notion inbox.
- Discuss with the agent. It asks clarifying questions, then writes out subtasks and dependencies on the row.
- Approve when you're ready. The task and its subtasks move to To-Do.
- Run. Agents pick up
Exec=Agentrows, work in parallel (respecting dependencies), and close them with result links.
That's the whole loop. The board is the source of truth; the agent is just an executor.
- Install agency-os in your harness (see the table below).
- Create a Notion integration at https://www.notion.so/my-integrations and share it with your workspace (or a specific parent page).
- Drop the token in
.env:NOTION_KEY=secret_...(or configure in your harness's MCP settings). - If using Cursor, Cline, Continue, or a generic agent: run
/agency-os initto configure which models to use for easy/med/hard tasks. (Claude Code handles this automatically.) - Run
/agency-os scaffold(or/agency-os scaffold --parent=<page-id>if the integration is scoped to a specific page).
That's it. Scaffold builds the Hub, Tasks database, corpus pages, and all linked views in Notion. No template to duplicate.
agency-os ships as a Claude Code plugin and as a portable skill spec for every other harness. The core contract (.claude/skills/agency-os/SKILL.md) is the same everywhere; only the wrapper changes.
| Harness | Format | Setup |
|---|---|---|
| Claude Code | Plugin | /plugin install https://github.com/ratamaha-git/agency-os |
| Cursor | Skill / rules | docs/harnesses/cursor.md |
| Cline / Continue | Custom instructions | docs/harnesses/cline.md |
| Any MCP-capable agent | Generic spec | docs/harnesses/generic.md |
All variants talk to Notion through the same MCP server, so your data and commands stay portable. Model selection (which models run which tasks) is configured via /agency-os init on non-Claude harnesses; Claude Code handles it automatically.
- docs/quickstart.md - 5-step setup with a worked example
- docs/concepts.md - status flow, dependencies, what lives where
- docs/extending.md - custom corpora, new commands, integrations
MIT - see LICENSE.
Built by AutomateLab.