Wire your agent

What you need

• Node.js 24 or newer. Not optional, and worth checking before anything else: on an older Node the install appears to succeed but silently gives you a placeholder release with no CLI in it. Run node -v first.
• A coding-agent harness: Claude Code, Cursor, or any other client that speaks MCP.

Nothing else.

Install and initialise

In your project directory:

node -v                # must print v24 or higher
npm install -g rcf-lite
rcf --version          # expect 0.15.0 or later
rcf init

If rcf --version comes back command not found, you are on an older Node and npm installed the placeholder release instead of the real one. The only tell is a single npm warn deprecated rcf-lite@0.0.1 line ahead of the cheerful added 1 package; easy to miss on a scrolling install. Install Node 24 (brew install node@24, or your platform's equivalent), then npm install -g rcf-lite again.

rcf init asks you for the project name and nothing else. Then restart your coding-agent session in that directory. The order matters: harnesses read .mcp.json and their instructions file at session start, so init has to have run first.

What rcf init just did

• Scaffolded the document chain under rcf/: placeholder documents for the product, its requirements, its architecture and its build queue, there for your agent to fill in.
• Registered the rcf MCP server in .mcp.json, so your agent gets the tools.
• Wrote a managed instructions block into both CLAUDE.md and AGENTS.md, so a harness that reads either picks up the operating rules.
• Seeded a knowledge space at rcf/knowledge/, the project's memory between sessions.

That is the whole survey. Init is safe to run in a repository that already has code: it touches only the paths above plus a managed block in .gitignore, and it never reads or writes your source. Adding RCF Lite to an existing repository sets out that contract in full and shows you how to prove it.

Ask it to build

Paste this into the session, with your own idea in the brackets:

AII want to build [describe your product idea in a sentence or two]. Let's get started.

What the agent will do next

First it checks the state of the project's documents and the work queue; that check is its habit at every session start, not something you ask for:

AIShow me the state of the document chain.

Terminalrcf define validate
rcf build queue

Then it walks its elicitation playbook with you: a short round of plain-language questions about the problem, who it is for, what done means, and where the thing will run, before it authors a single document. The chain is written from your answers, and you review it before anything gets built.

So this is what success looks like: your agent asking you questions, not drafting. If it is interviewing you about your product, the wiring took.

If something is off

• The harness did not pick up the wiring. It read its config before rcf init ran. Restart the session so it re-reads .mcp.json and the instructions file.
• The agent starts drafting the tree without asking you anything. The instructions block did not load. Run rcf doctor, which diagnoses and repairs wiring drift, then restart the session.

Where next

Driving is the supervision craft: what to look at, what to ask, and when to step in. And the rules your agent is now following are no secret: they are the managed block init wrote into CLAUDE.md and AGENTS.md, in your own repository, readable any time. If you want to read them here first, the harness block publishes the same text.