Morning briefing
Mail, meetings, calendar and PRs, read before you sit down and turned into proposals you can triage.
AI notes app · plan your day · work with agents
Ensemble is an AI notes app and shared workspace where you and your AI agents see the same picture: your notes, todos, meetings, mail, code and people. Plan your day, keep your context in one place, and let an agent do real work without being told the context every single time.
What it's for
Write notes, plan the day, keep context and hand off work, in one place your AI assistant and coding agents can read.
Notes, meeting notes and pages that your assistant and coding agents read as context.
Read moreMail, chat and meetings become proposed todos on one Today page, next to your calendar.
Read morePeople, projects and decisions kept once, then given to every LLM, agent and editor.
Read moreOne board, approvals, and a single place where every AI agent asks for you.
Read moreWhy
The context an agent needs already exists in your mail, chat, meeting notes, calendar, issues and repos. A chat box throws most of it away, then asks you to type it back in.
A day with it
Overnight mail and chat became proposed todos. Yesterday's design review is split into who owes what, linked to the sentence that said it.
Every proposal has three answers: I'll do it, Agent does it, Later. You decide. Nothing is assigned behind your back.
Agent tasks move across the board: gathering context, drafting, waiting for you. Ask for help in the editor and the context is already there.
A reply written in your tone and a PR description in your style for that repo. Approve, edit or reject. Forty seconds.
What shipped, what slipped, and a first plan for tomorrow. Every agent action sits in an audit timeline with what it read and who approved it.
What's inside
Gmail, Google Calendar, GitHub, Slack and Linear today, read-only. Outlook and Teams are next.
Mail, meetings, calendar and PRs, read before you sit down and turned into proposals you can triage.
You, Agent, Needs you. See what the agent is doing, how far it is, and what it plans next.
Sending, opening a PR, posting: anything that leaves the building waits for a yes from you.
People, projects, repos and preferences learned from your own work, so nobody re-types context.
Versioned skills for how you write PRs, commits and replies. Edit them like any other page.
A hash-chained, insert-only audit log. Every run is replayable, with its inputs and approvals.
An MCP bridge hands VS Code, Cursor and Copilot CLI the same briefs and skills, and can update your tasks when you allow it.
Gemini, OpenAI, Claude, OpenRouter, Ollama or Copilot. Keys are encrypted and never leave the runtime.
Needs you
For the tasks Ensemble hands them, Claude Code, Codex, GitHub Copilot and Cursor send their permission prompts and questions to Ensemble, with their own options and room for your own answer. Answer from the web, your phone or the terminal, or in the tool itself: in Claude both stay open and the first answer counts. Folders you trust in Ensemble are trusted in those tools too. The hook never allows on its own.
pnpm add -w p-retry@6CLI mode
Type ensemble in a project. One tab shows your runs and what needs you; the other is a chat with Claude Code, Codex or Cursor on your own sign-in, working in that folder with your Ensemble board loaded. Tasks you assign on the web run here too, in a sandbox the operating system enforces.
$ cd ~/code/shop-api && ensemble
[e] Ensemble Tasks · 1 needs you Chat
◆ Add a README too? Claude Code asks
1 Yes 2 ★ No 3 Skip or type an answer
› /run fix the flaky checkout test
Started Claude Code · Ask · may commit
› /profile
❯ Quick Ask · may commit · sandbox on
Careful Review only · local only
Overnight Run unattended · may pushCommands run inside the computer's own sandbox: Seatbelt on macOS, bubblewrap or Landlock on Linux, a separate sandbox account on Windows. No Docker, no VM. Ensemble sets it up; Windows asks once.
Hand a task to Claude Code, Codex (its CLI or the ChatGPT app) or Cursor's agent with your own sign-in, or to Ensemble's agent with a model key, under the same folder, git and internet rules. Or open it in the agent's own app or a terminal and press Send there. Ensemble never sees a sign-in.
Permission prompts and multiple-choice questions, with the same options and a box for your own answer, reach Needs me on the web, your phone, the app, or the terminal: click an answer, press its number, type your own, or /allow.
In a folder you trust, it works without asking, takes its recommended option at each decision, and lists every decision at the end. Outside the folder it may read, not write, unless you allow more; a refused path is never fatal.
Trustworthy by default
Run it on your machine as a CLI or a desktop app. They share one runtime, so install either first. No Docker at all.
Connectors read. Writes are separate actions with a preview and a policy.
Pause the agent, reject a draft, undo its writes. Control points are part of the design, not a setting.
Read every prompt in Settings. Read every line of the code that runs them, on GitHub.
What we measure against
Ensemble is early. These are the numbers it is built to move.
FAQ
Ensemble is an AI notes app and shared workspace where you and your AI agents work from the same context: notes, todos, meetings, mail, calendar, code and people. It plans your day with you, keeps your context in one place, and hands work to agents that ask before anything leaves.
Yes. Ensemble lives at ensemblework.com, the web app is at app.ensemblework.com, and the command-line tool is the ensemblework package on npm. People call it Ensemble or Ensemble Work.
Anyone who works with AI agents and LLMs. When you sign up you pick a role (student, teacher, lawyer, engineer, builder or manager) and a starting layout made for it. Engineers also get a CLI that runs agent tasks in a sandbox on their own computer.
Bring your own key for Gemini, OpenAI, Claude, OpenRouter, Ollama or GitHub Copilot. Ensemble can hand coding tasks to Claude Code with your own sign-in, permission prompts from Claude Code, Codex, GitHub Copilot and Cursor come to one place, and editors read your context over MCP.
Creating an account is free, and notes, tasks and context work without a model key. Model features use your own provider key, so the provider bills you directly. The code is source available on GitHub.
Yes. Use it in the browser, or install the CLI on macOS, Windows or Linux to run tasks locally in the operating system's own sandbox. No Docker or VM is needed.
Create an account, save your notes and context, and bring your own model key when you are ready to use an agent.