Onboarding into a new repo

You cloned a project and have 15 minutes to understand enough to start contributing. This workspace lets the agent give you the tour.

Nodes on the canvas

Node	Role
File Explorer	Points at the project root.
Code	Opens files the explorer or the agent clicks.
Terminal	For running npm install, pnpm i, cargo build, etc.
Agent Terminal	The Claude that'll investigate and narrate.
Text (named "Project summary")	Where the agent writes the conclusion.

Connections

From → To	Kind	What it does
File Explorer → Code	port ports:selectedFile->file	Click a file, it opens in the editor.
Agent Terminal → File Explorer	permission fs-full-access	Agent lists, opens, searches files.
Agent Terminal → Code	permission code-read-write	Agent reads and edits what it needs.
Agent Terminal → Terminal	permission execute	Agent runs commands.
Agent Terminal → Text	permission project-summary	Agent writes the summary straight into the text.

Permissions — what you're granting

• fs-full-access: tree, search, read — but no delete or rename.
• code-read-write: agent can make small changes (add comments, fix typos).
• execute: full terminal access — that's where the risk lives. Start with the repo on a clean branch.

Starter prompt

Give me a tour of this repo. Answer in three blocks:

1. WHAT: what this project does, in one sentence.
2. STACK: languages, main frameworks, and how it runs in dev.
3. ENTRYPOINTS: three files I should read first to start contributing,
   with one sentence per file explaining why.

After that, install dependencies and try running the app in dev mode.
If it breaks, tell me the error and where we are.

Write the final tour into the "Project summary" text node.

Why this setup works

• The File Explorer gives the agent a quick view of the tree without it having to fire 50 reads.
• Code with code-read-write lets the agent open files worth inspecting (package.json, README, main entry).
• Terminal with execute is where most of the work happens — install, dev, lint. Important: start on a clean branch.
• Text with project-summary is durable. You close the workspace, open it two days later, the summary is still there.

Variations

• Paranoia mode: swap code-read-write for code-read and execute for terminal-read. The agent only observes and writes the summary — touches nothing.
• Deep mode: add a Git Node wired to the Agent Terminal with git-log. The agent starts looking at recent commits to understand what's being built right now.