The Craftsman's
Manual

01Overview #

DAEDALUS is a coding agent powered by Claude. You give it a task in plain language; it reads your project, decides what to change, asks your permission, makes the change, and runs your tests to prove it worked.

It has three deliberate limits:

• Terminal only. No GUI, no browser tab, no editor plugin. If it isn't started in an interactive terminal, it exits.
• One folder. File tools can only touch the folder you launched it from and what's inside it.
• Asks first. Writing files and running commands need your approval, unless you turn that off.

02Requirements #

03Install #

One line. Pick whichever you like:

Try it without installing

npx daedalusrh

Install it for good

npm install -g daedalusrh

After a global install the short command daedalus works in any folder. Check it with daedalus --version.

Other package managers

pnpm dlx daedalusrh   # or: pnpm add -g daedalusrh
bunx daedalusrh       # or: bun add -g daedalusrh
yarn dlx daedalusrh

04API key & login #

DAEDALUS talks to Claude through the Anthropic API using your own account, so usage is billed to you. It picks up credentials in this order — the first one found wins:

1. ANTHROPIC_API_KEY environment variable
2. ANTHROPIC_AUTH_TOKEN environment variable
3. A profile saved by ant auth login

Option A — API key

$env:ANTHROPIC_API_KEY = "sk-ant-..."

setx ANTHROPIC_API_KEY "sk-ant-..."

export ANTHROPIC_API_KEY=sk-ant-...   # add to ~/.bashrc or ~/.zshrc to keep it

Option B — log in once

If you use Anthropic's ant CLI, run ant auth login once. DAEDALUS will use that profile automatically — no key to copy around.

05Your first task #

Go into a project folder and start the agent:

cd ~/my-app
daedalus

Or hand it a task straight away:

daedalus "add a /health endpoint and test it"

A typical turn looks like this:

  ─────────────────────────────────────── ◇ ask · ↑0 ↓0
  ❯ add a /health endpoint and test it

  ✦ thinking
  │ Read the server routes first, then add the endpoint and run the tests.

  ╭─ ◫ read_file  read src/server.js
  ╰─ ✓ 48 lines read · 5ms

  ╭─ ✎ edit_file  edit src/server.js
  │ -   "/": () => "hello",
  │ +   "/": () => "hello",
  │ +   "/health": () => JSON.stringify({ ok: true }),
  │ allow?  [y] yes  [n] no  [a] always  › y
  ╰─ ✓ edited src/server.js · 4ms

  ╭─ ❯ run_command  $ npm test
  │ allow?  [y] yes  [n] no  [a] always  › y
  │ 12 passing
  ╰─ ✓ exit 0 · 2.1s

  Added GET /health returning { ok: true }. All 12 tests pass.

  ✻ forged in 14.2s · ↑3.3K ↓530 tokens · 5.0K from cache

06Reading the interface #

Replies are rendered as light markdown: headings, bullet and numbered lists, bold, inline code and boxed code blocks.

07Approvals #

Reading is free. Changing things needs a yes. When a tool needs approval you'll see a preview first — a red/green diff for edits, the first lines of the file for writes, the exact command for shell runs — then:

08Commands & keys #

The conversation is remembered across turns until you /clear or quit, so you can say things like "now add a test for that too".

09Writing good tasks #

• Say what "done" looks like. "Fix the login bug — the tests in auth.test.js should pass" beats "fix login".
• Name files when you know them. It can search, but a pointer saves time and tokens.
• One task at a time. Big rewrites go better as a few smaller turns.
• Any language works. It replies in the language you write in.
• Interrupt freely. If the thinking goes the wrong way, press Ctrl+C and steer.

10Tools #

The agent can do exactly six things. Nothing else is available to it.

11Configuration #

12Model & limits #

• Model: Claude Opus 5.5 by default, with adaptive thinking at high effort.
• Reply length: up to 64,000 output tokens per reply.
• Fallback: if the main model declines a request, the API can retry it on another Claude model automatically.
• Caching: repeated context is cached, which makes long sessions faster and cheaper. /cost shows the cache reads.
• Memory: only the current session. Nothing is remembered after you quit.

13Safety model #

• Workspace lock: every file path is resolved against the launch folder; anything that points outside (../, absolute paths) is refused.
• Approval gate: writes, edits and commands wait for you unless auto mode is on.
• Visible changes: you see the diff or file preview before approving.
• Timeouts: shell commands are killed after 120 seconds.

14Privacy & data #

To reason about your code, DAEDALUS sends these to the Anthropic API as part of the conversation:

• what you type,
• the contents of files it reads and the results of searches,
• the output of commands it runs,
• your workspace path and operating system name.

Nothing is sent anywhere else, and DAEDALUS keeps no logs on disk. Files it never reads are never sent. Anthropic's handling of API data is described in their privacy policy and commercial terms.

15Cost #

DAEDALUS itself is free and MIT-licensed. You pay Anthropic for the tokens it uses, at the normal API price of the model you've chosen. To keep costs down:

• run it inside the specific project, not a huge parent folder;
• /clear between unrelated tasks so old context isn't resent;
• check /cost now and then;
• set a monthly spend limit in the Anthropic Console.

16Troubleshooting #

17Uninstall #

npm uninstall -g daedalusrh

That's it — DAEDALUS stores nothing else on your machine. If you only ever used npx, there's nothing to uninstall.