
Pia · Planner
Opus · high
Plans the big or unclear tasks on Opus: files, steps, how to check. Never edits.
quack:planner
Quack tells you when to /compact or /clear, maps your code so Claude reads only the lines it needs, and hands the work to a team of five ducks, each on the right model for its job.
> /plugin marketplace add AlekDob/quack-marketplace> /plugin install quack@quack-marketplaceThen restart Claude Code. Once per project: /quack:index and /quack:checkup.
The duck team
Five ducks, each a subagent with its own model and effort. Opus plans and codes when it matters; cheaper models search, research and take notes. Only their summary comes back to your chat.

Opus · high
Plans the big or unclear tasks on Opus: files, steps, how to check. Never edits.
quack:planner

Opus · medium
Builds the plan or a long job on Opus, in his own context, runs the tests, returns only what changed.
quack:coder

Haiku · low
Finds where things are. Paths and line ranges, never file dumps.
quack:scout

Sonnet · medium
Looks things up on the web and answers with sources.
quack:researcher

Haiku · low
Writes the diary, gotchas and decisions: the memory the next chat starts from.
quack:scribe
What we measured: the same multi-file task cost $0.25 on Sonnet alone, $0.56 on Opus alone, and $0.63 with a Sonnet chat handing it to Pia and Kit. The ducks buy quality on hard tasks, not a cheaper bill. The cheap lever is your chat's model. You can always ask for a duck by name: "have Pia plan this".
Why chats get expensive
A chat at 150k tokens pays to read 150k tokens on every message, even for "ok, thanks". The prompt cache makes that cheaper, but it expires after 5 minutes on the API and 1 hour on subscriptions.
Bug fix, new feature, quick question: all in the same chat. Every new task carries the weight of the old ones.
Claude opens a 1,600-line file to change ten lines, and the whole file stays in the chat.
Searching code or the web on the most expensive model, at the highest effort, costs far more than it needs to.
Long memory files, dozens of skills and MCP servers are paid again in every single message.
What Quack does
Everything runs locally with Node, with no dependencies. About 750 tokens per session.
At every start Claude gets about 330 tokens of rules: short answers, read ranges instead of whole files, one chat per task, the right duck for each job.
Once every 100k tokens, and after a pause that expired the cache: "New task? /clear. Same task? /compact."
Before Claude reads a file of 400+ lines whole, Quack points it to the code map or Grep, then to a line range.
/quack:index writes documentation/INDEX.md: folder, file, purpose, symbols with line ranges. It refreshes itself every session.
Pia plans and Kit codes on Opus; Marco, Iris and Dante search, research and write on cheaper models. Each sends back only the summary.
Diary, gotchas and decisions live in documentation/. A new chat gets the last diary lines, so /clear is not starting over.
How it works
No new app to learn: Quack plugs into the Claude Code you already use.
They run on their own at the right moment: when a session starts, when you send a message, before a file is read. They never block your work.
/quack:index builds the code map. /quack:checkup measures what your setup loads into every message and fixes it, only with your OK.
The five ducks each have their own model and effort. Claude hands them the work by size; the expensive model keeps only the result.
What no plugin can do: run /clear or /compact for you, or switch the main chat's model. Quack tells you when; you type the command.
Free guide
Why Claude Code burns tokens, the habits that fix it, and how Quack automates them. Read it online or download the PDF.
Read the code, open an issue, make it better. Quack is built by Alek Dobrohotov and Studio Futuro.
Quack started as a desktop app for running many AI agents. The part that mattered most was spending fewer tokens, so that is what Quack is now. The app is retired; its source code stays on GitHub.
Desktop app sourceA free plugin for Claude Code that cuts token waste. It tells you when to /compact or /clear, maps your code so Claude reads only what it needs, and hands searches, research and docs to cheaper models.
In Claude Code, run /plugin marketplace add AlekDob/quack-marketplace, then /plugin install quack@quack-marketplace, and restart. Then run /quack:index and /quack:checkup once per project.
No. Quack is free and MIT licensed. It adds about 750 tokens per session, usually less than a single whole-file read it prevents.
No. Everything runs locally with Node, with no dependencies. The only network call is one anonymous ping when you install or update Quack: the version and a random id, nothing else. QUACK_NO_PING=1 turns it off.
No plugin can: Claude Code does not allow it. Quack tells you when it is time. The automatic switch happens through the duck team: each duck has its own model and effort, and Claude picks the duck by the size of the task.
We retired it. Quack is now only the plugin. The app source code stays on GitHub.
Two commands in Claude Code. Free and open source.
> /plugin marketplace add AlekDob/quack-marketplace> /plugin install quack@quack-marketplaceHave an idea? Write us: quack@quack.build
Free and open source, made by indie developers in Puglia, Italy.
Not Silicon Valley. Puglia, Italy.