Docs

Building it

Coding agents

A coding agent such as Claude Code or Cursor can build straight from the idea. It reads the idea from Walden as it works, keeps its list of tasks here, and asks you when the idea doesn’t answer something. It can’t change a word of the idea.

How it goes

  1. Make a key and give it to your agent. You do this once.
  2. Hand an idea off with With a coding agent.
  3. Open your agent in the repository where the idea will be built and ask it to build the idea. Then follow the build from the idea, answer what it asks, and review what it made. See Following a build.

Make a key

Go to Settings → Coding agents and choose New key. If you have no key when you first hand an idea to a coding agent, Walden makes one for you.

  • A key opens every project in its workspace, including ones made later, and the ideas that aren’t in a project. It opens nothing in another workspace: make a key there for that.
  • It acts as you, and never has more than you have. It stops working if you leave the workspace or delete your account.
  • Anyone in the workspace can make a key, viewers too, since the engineers who build are often viewers. Handing an idea off still takes an owner or an editor.

Connect Claude Code or Cursor

A new key is shown with what to give your agent, on two tabs, with Copy beside it. Walden’s address and the key are already filled in. Command is for Claude Code. Run it once in a terminal:

claude mcp add --transport http walden <Walden’s address> --header "Authorization: Bearer <your key>"

Settings file is for an agent set up with a file, such as Cursor. Add these lines where its settings list its servers (in Cursor, inside mcpServers in mcp.json):

"walden": {
  "type": "http", "url": "<Walden’s address>",
  "headers": { "Authorization": "Bearer <your key>" }
}

Hand an idea to it

  1. Open the idea’s More actions menu (…) and choose Build it….
  2. Choose With a coding agent and press Hand off. If things are still open in the idea, the button says how many, such as Hand off with 2 open. See Handing off.

Walden then gives you one line to pass to your agent, with Copy:

Build the Walden idea <the idea’s id>. Read it with get_idea before you start.

If you already have a key, it opens this idea too, and Walden says so: “Your key ending 7Qx2 opens it. Nothing to paste again.”

Start the build

Open your agent in the repository where the idea will be built and give it the line. Or just ask it what there is to build. It lists the ideas handed off to a coding agent, and where your agent can show you a list, you choose there: the ideas of the project your repository belongs to come first, with a row for the ideas in other projects. Tick one, then Accept. Where it can’t show a list, the agent asks you in words. In Claude Code, the slash command /mcp__walden__build does the same: give it an idea’s name, part of it or its id, or leave it out to choose from what’s waiting.

From there the agent reads the idea, breaks the work into tasks and starts. When the project has more than one repository, the agent is told which one it’s in and asked to open the others in the same session. See Repositories.

What the agent can do

  • Read the idea as it stands now, with its other pages, its UI sketches and the text of its files. An open unknown comes marked as not decided.
  • Look things up before a product decision: in this idea, and in what earlier ideas ruled out or settled. See Decisions from earlier ideas.
  • Break the idea into tasks, once, each named after the part of the idea it builds, and say when it starts and finishes each one.
  • Ask a question the idea doesn’t answer, and record the answer you give it where it runs.
  • Say it’s done, with what went differently from the idea and the address of its pull request.

Coding agents don’t use your credits. What the agent does with what it reads is between you and whoever makes it.

What it can never do

  • Change a word of the idea or of any of its pages. The answer to a question reaches the page only when a person puts it there.
  • Mark the build built, or change a task a person has taken. Those are a person’s.
  • Read an idea that wasn’t handed to a coding agent, or anything outside its key’s workspace. An idea you reopen is closed to it until you hand it off again.

One person’s agent at a time

The first agent to choose an idea from the list, or to make its list of tasks, picks it up. While it has the idea, another person’s agent can read it but change nothing, and is told who has it. Your own other keys carry on as before. Taking it over is your choice: where your agent can show a list, it asks you a second question, Accept to take it over or Decline to leave it; otherwise the agent asks you in words. Once taken over, the other agent can’t work on it, and the tasks it made are kept. If the person who has it takes their key back or leaves the workspace, anyone’s agent can pick it up.

Take a key back, or turn agents off

In Settings → Coding agents, choose Take it back beside the key. The agent it was given to can open nothing from then on, and whatever it had picked up is free for anyone’s agent. In a team, owners also see Other people’s keys and can take any of them back. If a team’s plan ends, its keys open nothing until it subscribes again.

An owner can untick Coding agents can read this workspace’s ideas in Settings → Coding agents. Off, every key in the workspace opens nothing, and Build it has only the file. Tick it again and the keys work as before.