Any OpenAI-compatible model
Point it at a base URL, pick a model from the list the server returns, and optionally give a small fast model the side jobs: session titles, context compaction, tool preselection and follow-up suggestions.
min-agent
A very small self-hosted agent: chat over any OpenAI-compatible server, with MCP tools.
What it does
A chat app and tool-calling loop for a model you already have: one you run yourself with Ollama or LM Studio, or one behind an API such as OpenAI or OpenRouter. It connects to MCP servers for tools, resources and prompts, and keeps its settings, servers, sessions and every message in Postgres. One front end ships to the browser, Android, and a Windows or Linux desktop app.
-
-
MCP tools, loaded on demand
Stdio and HTTP MCP servers plug in from the settings screen, and by default the model gets a name-only catalogue and loads the tool definitions it needs instead of being sent every schema on every request.
-
What every turn cost
Each reply reports its output tokens, tokens per second, time to first token and total time, beside a context meter and a breakdown of what filled the window.
Get started
git clone https://github.com/cubicecho/min-agent
cd min-agent
docker compose up -d
That builds the image and starts it with its Postgres database. Open
http://localhost:8787, go to Settings → Agent, enter the base
URL of an OpenAI-compatible server, refresh the model list, pick a model
and save.
Nothing in min-agent is authenticated, so publish the port only on a network you trust.
Full documentation lives in the project README: https://github.com/cubicecho/min-agent