Consensus-AI

Plans you and your LLM actually agree on.

Consensus-AI turns a draft into a plan your LLM can build, one settled question at a time.

A plan open in Consensus, with its threads beside the text. A plan open in Consensus, with its threads beside the text.

You have had this conversation before.

Working a design out with an LLM is fast, until it is not.

Decisions buried in a chat scroll

You settled that question an hour ago. Now it is forty messages up, and the model has forgotten it.

The wrong thing, built confidently

One sentence read two ways. The model picked one, built it beautifully, and now you are unpicking it.

No record of what was agreed, or why

The plan that was built is not the plan you have. The reasons are somewhere in a conversation nobody will reread.

How it works

A plan is a document with threads. Each thread is one question, settled in one place, and applied to the text in front of you.

1

Write a draft

Markdown, in an editor with a live preview. Rough is fine: the draft is where the questions come from, not the answers.

The editor beside the live preview. The editor beside the live preview.
2

Settle it thread by thread

Select a passage and comment. The LLM replies. You resolve. It applies the outcome to the text, and the thread turns green with the revision it made.

The threads pane: contested, resolved and applied threads. The threads pane: contested, resolved and applied threads.
3

Build it

The plan is done when every thread is applied. Say "build it": your agent implements exactly that plan, and its build log is a thread too, accepted the same way.

A build log thread, applied on a finished plan. A build log thread, applied on a finished plan.

What you get

A record you can trust, and an LLM that only does what you settled.

Where it runs

On your machine, on a server of yours, or in a window of its own.

The browser

A small Node server and a page. Start it, open it, and you are the owner.

The desktop

The same page in its own window, connected to any server or running one locally, with LLM connections and workspaces of its own.

Docker

One image, one volume, port 3500. Put it on a box on your network and work from anywhere on it.

The desktop's connect screen. The desktop's connect screen.

Get started

The code is public at github.com/mchiver/consensus-ai. Four ways to run it.

From the checkout

Node 22 and git

Clone, install, start, and open the page.

git clone https://github.com/mchiver/consensus-ai.git
cd consensus-ai
npm install
node bin/consensus.js

Then open http://127.0.0.1:3500/. More in the guide.

In Docker

One image, one volume

Build the image from the checkout; the data lives in its volume.

docker build -t consensus:latest .
docker compose up -d

Port 3500. More in the guide.

The desktop

A window of its own

Run the desktop from the checkout and connect, or start a local server from its connect screen.

npm run desktop

More in the guide.

The installer

Windows, no checkout

The desktop packaged with its server: download, install, open. The installer is not signed yet, so Windows asks once before it runs.

Download the latest release

More in the guide.

Consensus-AI is not Consensus, the research search engine at consensus.app, and has no connection to it. Consensus-AI is a tool for settling a plan with your LLM, thread by thread.