Skip to main content
Milford runs flows: graphs of small steps such as typed decisions, model calls, templates and HTTP calls. You configure it as code, run it as a TypeScript library or a single Docker container, and call it over HTTP. Milford is not an agent framework. There are no agent loops and no tool-use planning. A flow is a fixed graph that you can read, version in git and test.

​
What you get

  • A core library, @milfordai/core, with no I/O and no vendor knowledge.
  • Provider adapters in @milfordai/providers: OpenAI (and compatible servers), Anthropic, Jev and generic HTTP.
  • A thin HTTP server, @milfordai/server, with bearer auth and server-sent events.
  • An MCP server, @milfordai/mcp, that exposes flows as tools to external LLMs.
  • Docker images for linux/amd64 and linux/arm64. No database.

​
Where to go next

Installation

Install with npm or pnpm, or run the Docker image.

Quickstart

Run your first flow.

Flows

How the executor walks a graph.

Decisions

Typed answers with probabilities.

Home automation

A free-text command turned into device actions.