Claude Code · multi-agent workflow

Fable where it counts.
No bills of mythical proportions.

Purser is a Claude Code workflow that writes and ships code for you. It uses the expensive models, Fable and Opus, only where they earn it, like planning and review, and cheaper models everywhere else. You see what a task will cost before it runs.

$curl -fsSL https://get-purser.pages.dev/install.sh | sh

Requires Claude Code. Installs one workflow file into ~/.claude/workflows/, no clone, no npm. Then type /purser in any session.

The pipeline

Eight phases. The cheapest model that does each job.

Casting is deliberate: mechanical steps run on fast, cheap models, and the top tier goes only to the work that needs it: planning, hard changes, and the adversarial review.

0 · 1

Preflight & Explore

Repo sanity, default-branch and verify-command detection, per-repo lock. Read-only scouts map the task surface and blast radius.

Plan

The one stage always worth the top tier. Emits file-disjoint work items with per-item fast-verify commands.

Implement

One fresh git worktree per item, in parallel. Implement, verify, commit, publish a branch. Retries escalate to a stronger model.

Integrate

Merge the item branches and run your full suite once. A red build gets one repair pass.

Review

A top-tier model adversarially reviews the diff, read-only. It fixes serious findings; a green build is never left unreviewed.

Finalize

push --force-with-lease, open the PR, clean up branches, release the lock. You review one PR.

Spend, on your terms

Top-tier model spend is never a surprise.

Purser prices each work item and lets you decide who pays for the premium tier, before any money moves.

Preview, don't guessA quote prints each item's plan and a dollar estimate. Nothing is built until you approve.
A disinterested judgeThe tier selector sits outside the models it chooses between, so it has no stake in recommending itself.
Every run is loggedThe PR body records what was decided, by whom, and what it was estimated to cost.
Roughly half, and you see your own numberIn our own runs, keeping Fable for planning held spend to about half of running the whole task on the top tier. Purser shows you that split for your task before you approve. It is a live quote, not a marketing number.

Modes

One tool, four gears.

All opt-in. Stack them freely, for example --quote --explain.

--quote

Preview & price

Plan the work, print the per-item recommendation and a cost estimate, and save the quote without building anything. Approve, then --execute.

--innovate

Spend up where it counts

A disinterested selector escalates the hard, high-blast-radius items to the premium tier within a dollar budget you set.

--explain

Learn as you build

A tutor explains the system design of the feature being built, its components, tradeoffs, and the CS/SWE concepts, so real work doubles as interview practice.

--audit

Was it worth it?

A skeptical, disinterested auditor rules whether each premium-tier item actually earned its cost, and can say it didn't.

Install

One line. No clone.