Skip to content
mercury framework
GitHub

Mercury Framework

Mercury Framework (Mercury, for short) is a framework for building your own agent: a program that talks with people in a chat, understands what they ask with a language model, and does it with the tools you give it, such as searching the web or opening a ticket.

You describe the agent in one file: what it's called, where people can talk to it, what it's allowed to do. Mercury takes care of the rest, from the conversation to the memory of what was said, and it runs on your own machine or server, in Docker.

What you get

Channels
Where people talk to the agent: Google Chat, or an HTTP interface for a chat of your own. A terminal is always there for you.
Plugins
What the agent can do. Each plugin adds tools (Jira, Bitbucket, web search…) and the agent can use nothing it wasn't given.
Confirmations
An action that can't be undone never runs because the model decided so: the person gets a one-time code and the action runs only when they send it back.
Memory
The agent remembers the conversation, keeps notes in a wiki, and recalls earlier conversations. Each person has their own.

What it isn't

It isn't a service you sign up for, and it doesn't come with a language model. The model is yours: Mercury talks to any server that speaks Ollama's protocol, on your laptop or on a machine in your network, so the conversations never leave your hands.

What you need

  • Bun, the JavaScript runtime Mercury is written for.
  • Docker, where the agent runs.
  • An Ollama server with a model that can call tools, reachable from Docker.

Build your first agentfrom an empty folder to the first answer