Todd.

what it is.

You describe an outcome, like “launch a waitlist site for dropin.hockey with Firebase auth.” Todd’s planner decides which agents the job needs, gives each one the right tools (a code sandbox, a real signed-in browser, infrastructure APIs, plugins and MCP servers) and runs them in parallel. It pauses for approval before it spends money or posts anything.

highlights.

• No hardcoded team: the planner spawns agents on the fly with the toolsets each part of the job needs, up to six at once by default.
• Agents never collide: one agent owns each piece of code, and Todd refuses edits to a file another running agent is working on.
• Every agent has a live window with its reasoning, tool calls and browser steps, plus pause, resume and steer controls.
• Sign in once: logins in the agents’ browser are detected, and the matching CLIs (GitHub, Vercel, Firebase, Expo and more) are signed in with the same session.
• Runs end with a structured summary per agent, and a finished run can be continued with a follow-up message.

how it's built.

Python at the core: a FastAPI service whose LangGraph planner, checkpointed in Postgres, spawns and coordinates the agents. Agents get toolsets: a sandbox container for code and deploy CLIs, a headful Chromium driven by browser-use with a live noVNC view, infrastructure APIs and MCP servers. A Next.js dashboard streams it all over server-sent events. Secrets live in an encrypted vault and never enter model context, and the sandbox and browser are network-isolated. Docker Compose runs everything, and models are bring-your-own. NeuralDaily and Rill were both built with it.

• Python
• FastAPI
• LangGraph
• Next.js
• MCP
• Docker

latest activity.