Missions

Missions is a desktop app that gives a founder an AI team to build and maintain their app — without asking them to read code. Underneath it runs our Agentic Workflow, the same process we use to plan, build, review and ship our own products, on Claude Code.

The founder starts a mission from a short form — What should we work on? — answers each decision the workflow puts to them, previews the result, and presses Release. Nothing ships without that press; the agents can propose a release, but only the owner can confirm one.

Everything runs on the founder's own computer, on their own Claude plan. There's no separate API bill, and Missions never touches their credentials — sign-in happens through Anthropic's and GitHub's own flows.

AI agents can now ship real software, but every tool that drives them assumes the person in the chair is a developer. Terminals, diffs and raw prompts are the interface — and the decisions a founder actually owns (what to build, whether the plan makes sense, whether the result is ready to ship) are buried inside them.

The brief was to keep the engineering discipline and change the interface. Scope approval, plan review and release stay exactly as strict as they are in our own work. What changes is how they arrive: each one reaches the founder as a plain-language question, with the real options in front of them.

Missions is a TypeScript monorepo around an Electron shell. The app drives Claude Code on the founder's machine and talks to it over a localhost-only, token-protected connection — nothing is exposed beyond the computer it runs on.

The workflow itself is used unmodified, at a pinned version. Every new version has to pass a suite of contract tests before Missions accepts it, and founders can check for updates, switch versions, or roll back if something feels wrong.

Decision points are the heart of it. When the workflow needs an answer, the session pauses and the question appears in the app as a card. The founder answers, and the work resumes exactly where it stopped. Every step, question and answer is written to an append-only log, so there's always a complete record of what happened and why.

The app follows a shape anyone who has used a chat tool will recognise. Projects sit in the sidebar with their missions underneath — one line and a status dot each. A mission opens as a conversation, with the detailed work folded into steps, and a composer for talking to the team directly. A side panel holds the plan, a live preview, the release action and the project's docs. Keyboard shortcuts cover the common moves, and the model and tools each agent uses can be tuned and saved per project.

The desktop app is available to download now as a public beta. Next is a hosted version: missions keep running while the founder's laptop is closed, and decisions can be answered from a dashboard or straight from their inbox.

If you're a founder with a product to build and no engineering team behind them, download it and start your first mission. And if a question ever stops making sense, tell us — that's the feedback we're building around.