o1-spec/foundry

Foundry is a hackathon project that gives you a visual command center for running teams of AI agents on a single task. Instead of one AI model handling a request start to finish, Foundry breaks the work into stages and hands each stage to a specialist agent, then lets you watch, pause, and steer the whole process from a browser interface. The README frames the problem this way: building systems with multiple cooperating AI agents is difficult because context has to be passed correctly between steps, token costs can spiral, latency bottlenecks are hard to trace, and there is often no way for a person to step in before something goes wrong. Foundry addresses this with a visual execution graph built on React Flow, so each agent shows its current status such as waiting, running, retrying, completed, or failed. A single run deploys seven specialist agents through a mix of sequential and parallel stages. A Planner creates the task roadmap first. A Researcher and a Product Manager then work in parallel to gather information and write requirements. An Architect designs the API and database structure, an Engineer produces mock code and layouts, a QA Engineer writes test checklists, and a Lead Reviewer combines everything into one final package. Throughout a run, a person can inject new instructions, ask a specific agent to revise its output, regenerate a single artifact without rerunning the whole pipeline, and give final approval before the package is considered done. Under the hood, Foundry talks to an OpenAI-compatible endpoint called BTL Runtime, reading custom response headers to track cache hits and estimated cost savings for each call, and it streams live reasoning and progress updates to the browser over server-sent events. It is built with Next.js and TypeScript on the frontend and backend, PostgreSQL with Prisma as the database layer, and Tailwind CSS with Shadcn UI components for styling. Running it locally requires cloning the repository, installing dependencies, configuring a .env.local file with a database connection string and a BTL Runtime API key, initializing a local PostgreSQL cluster, and seeding the database before starting the development server. The README does not state a license.