Meridian sits between your clients and an OpenAI-compatible inference endpoint. To a client it looks like an inference endpoint, but your auth (workOS) carries you: your agents, your threads, and whatever your extensions add.

• /v1 is an OpenAI-compatible proxy. Point any OpenAI client at it and pick an agent as the model.
• /api keeps threads on the server. A reply keeps generating if the client disconnects, and the client can pick the stream up where it left off.
• Extensions are sandboxed Lua scripts, edited in the browser. They hook into every run and can add tools. Basically anything programatic with a similar philosophy as the Pi harness
• Skills as normal skills, nothing special, they can recommend extensions.

It is built with Elixir, Phoenix and Postgres. Every run, tool call and script call is its own process, so something that hangs or crashes takes down only itself. The admin console shows usage and run metadata, never the conversations.

Early development. Container images are published for amd64 and arm64.