Run it on your next engagement
Bring the agent you already use. The record stays on your machine.
YarraMate is a tool your agent calls, not a chat you log into. It runs where your files are, so the agent has to be able to reach them. Most agents can. The ones that cannot reach files use a hosted workspace instead.
Supported today
- Claude Code, in the terminal or your editor
- Codex CLI, in the terminal
- Cursor, in the terminal or your editor
- Claude Desktop app
- Codex desktop app
All five run on your own machine and can read and write the record in your repository. Nothing is sent to us.
Through a hosted workspace
- claude.ai in a browser or on a phone
- ChatGPT in a browser or on a phone
Those agents cannot reach files on your computer, so the record lives in a workspace on yarramate.dev and they connect to it over the network. A$5 per seat per month, readers free, the first week free: pricing. Or use the desktop app from the same vendor, on the same subscription, and keep the record on your machine for free.
Setup
Once in a terminal, then one line for your agent.
Needs Node.js 22 or newer. Run the first command in the repository that should hold the record.
npx yarramate init .
claude mcp add yarramate -- npx -y -p yarramate yarramate-mcpor, as a plugin
/plugin marketplace add yarrasys/yarramate /plugin install yarramate-architecture@yarramate
codex mcp add yarramate -- npx -y -p yarramate yarramate-mcp
{ "mcpServers": { "yarramate": { "command": "npx",
"args": ["-y", "-p", "yarramate", "yarramate-mcp"] } } }{ "mcpServers": { "yarramate": { "command": "npx",
"args": ["-y", "-p", "yarramate", "yarramate-mcp"] } } }Then tell your agent where the record is. Every tool takes a workspace path; the desktop apps need the full path to .yarramate/workspace.yaml.
What happens next
The same loop the home page shows, on your own system.
The agent reads the brief and puts the first open question to you, with the reason it matters.
The agent turns the sentence into a change to the record and shows it to you before it lands.
The record is validated: every relationship allowed by the vocabulary, every view still drawable, the open-question count updated.
The documents are regenerated from the record. Commit them with the code; a pull request is the approval.
The one thing the browser cannot show
Checking the record against real code.
There is no code in a browser. Once the record lives in your repository, the same tool reports where the code and the record disagree, on every pull request. A worked example on a public codebase is in the builders' showcases on GitHub.
Prefer the editor?
The map you tried is one request away.
Ask your agent to open the live map beside the conversation, and watch the answers land on it as they are made. Right-click any subject for its open questions; every row can be answered on the map or handed back to the agent.