Docs / Overview

Riverforge documentation

Riverforge is a coding agent that runs entirely on your own machine. It lives in VS Code: chat, plan, edit, run tests, approve and roll back — with a memory that grows and an identity it earns. No cloud, no keys, nothing leaves your PC. This is the manual for installing it and getting real work done.

What Riverforge is

Riverforge is a fully local coding agent. You describe a change in plain English — “add retry logic and a test” — and it plans the work, reads and edits your files, runs the tests, shows you the diff, and remembers what it learned. The whole agentic loop happens on your own GPU through Ollama. There is no server to send anything to, and you can prove it with netstat.

The product is the workflow, not the model. A small model wrapped in a serious workflow — it inspects your code, makes the smallest safe change, runs the tests and remembers what it learned — behaves like a much larger agent, on hardware most developers already own, for free, with nothing leaving the box.

The pieces

Riverforge is a local server with several thin clients in front of it. You mostly live in the VS Code chat; the tray keeps the engine running; the visualiser is for tending the memory.

New here? Start at the top

If you have never run Riverforge before, follow these three pages in order. Twenty minutes from here you will be pair-programming with a fully private agent.

1. Installation — requirements, the Windows installer, Ollama and the model download, and getting the VS Code extension in place.
2. Quick Start — start the tray, open the chat, send your first request, read the diff and run the tests.
3. Chat Modes & Controls — Chat vs Ask vs Agent, the composer controls, attachments and the thinking toggle.

How a turn flows

Behind a single chat send, Riverforge moves through the same few steps. It’s a handy mental model — most of these docs are just a closer look at one of these stages.

• Recall — relevant memory about you, this project and your research grounds the request first.
• Plan — your message is turned into a task and handled as Chat, Ask or Agent.
• Edit — in Agent mode it reads the files, makes the smallest safe change and shows you a diff.
• Verify — changed files can’t be called “done” until your tests or lint actually pass.
• Remember — useful, durable facts are saved to memory in the background.

Find your way around