michaelshimeles/adam

Adam takes Vercel's eve, a durable AI agent runtime that normally runs as a long running Node server backed by Postgres and Redis, and rebuilds it so that everything runs on Convex instead. This is not just moving where the data is stored, it moves the actual execution engine itself. The eve build tool is used like a compiler, and the resulting server bundle is loaded into a Convex deployment and run inside Convex actions, so at runtime there is no separate eve server running at all. One Convex deployment ends up holding everything the system needs: the agent runtime itself, the workflow state such as runs, steps, events, and hooks, a message queue with retries and backoff handled by Convex's own scheduler, live streaming output from the AI model that the dashboard can watch update in real time, cron based schedules translated from eve's markdown schedule files, ordinary application data like a demo notepad tool, and even the dashboard web app itself, which is a Svelte 5 interface served directly from Convex. The repository is organized into a few main pieces. The backend package holds the Convex deployment along with the vendored eve bundle. A separate package implements the storage, queue, and streaming interfaces that eve expects, built specifically for Convex. An agent package defines the actual AI agent, its tools, and its schedule, and gets compiled with the eve build tool rather than run as a live server. A web package holds the Svelte based dashboard for chatting with the agent and watching its activity. There is also a platform piece that lets someone configure a new agent through a dashboard and deploy it to its own Convex project with one click. Getting it running locally involves installing dependencies with pnpm, starting a local Convex development deployment, and setting a handful of environment variables such as a world service secret, an execution mode flag, the Convex URL, and credentials for whichever AI model the agent will call. The full README is longer than what was shown.