Project Architect turns Claude Code into a small team that works from a plan. It is a project management system and a token-saving system in one. As a project manager, it keeps the plan: phases with a milestone a machine can check, your approval before any work starts, and a memory in your repository that every new session starts from. As a token saver, it runs each task in a small, fresh context, and helpers do the long reading and the build-and-test loops, so your session never carries them; a local ledger measures what that saves, and the statusline shows it. What PA3 is and does explains every part in plain words.

What you get

• A plan that outlives every session. Phases with a milestone a machine can check, task summaries, decisions, rules and techniques, all kept as files in your repository. A new session starts from them, not from a transcript.
• Work that runs on its own between two gates. You approve the plan and confirm the close; in between, nobody stops to ask unless a decision is genuinely yours.
• Savings you can see. A local ledger prices every request, and the statusline shows, live, what the helpers kept out of your context.
• Models that fit your plan. Max 20x, Max 5x or Pro: PA3 reads which from your login and picks its models to match.

See it

White is what you used; blue is what the helpers saved. Here this session used 1 % of the five-hour window and its helpers saved another 2 %, and over the project's life its work has gone 2.6 times as far as it would have without them. Every number, line by line.

Install

You need Claude Code 2.1.292 or newer, a Claude subscription (Max 20x, Max 5x or Pro), Python 3.12 or newer, and git.

1. Once per machine. Download this repository (Code → Download ZIP, then extract) and open its project-architect-3.0 folder. On Windows, double-click install.cmd; on WSL, Linux or macOS, run ./install.sh. It asks nothing, and it also installs the mod, the Claude Code plugin that runs a plan's tasks.
2. Once per repository. Open the repository in Claude Code: PA3 shows one line that sets it up, with the path already filled in. Type it, then restart Claude Code (/exit, then claude). Or double-click setup-project.cmd and pick the folder. It makes one commit and never pushes.
3. Start. Open a bare claude in your repository and talk with Claude, who works out where the project stands and takes it from there. The first session shows a line pointing to the Claude app's Settings → Billing page and the command that stores the day your plan renews, so the month figures count from it.

Step by step, updating and removing: Install. Every switch: INSTALL.md.

How it works

flowchart LR
    you((You)) <--> claude["Claude"]
    claude --> thinkers["thinkers"]
    claude --> retrievers["retrievers"]
    mod["the mod<br/>runs the loop"] --> expert["expert<br/>one per task"]
    expert --> coder["coder"]
    expert --> retrievers

Loading

You talk with one partner, Claude, in one session. Claude drafts each phase's plan with you; after you approve it, the mod runs one expert per task in a fresh context while you keep talking with Claude, and the expert briefs coders and retrievers. Every change an expert wants to make to the plan goes to Claude, who settles what stays inside the milestone and asks you about the rest. A task that hits a wall comes back blocked with its evidence and a recommendation. When the milestone is met, Claude asks you before the phase closes, and the close leaves a plain-English recap of what was done. What PA3 is and does · The roles · A day in the life

Why Project Architect

A long Claude Code session re-sends everything it has seen with every request: the files it read, the command output, its own reasoning. The context keeps growing, and every later request pays for all of it again. Project Architect changes the shape of the work:

• Memory that carries forward. Every session, and every expert inside one, starts from the files the work left behind (the plan, the task summaries, the decisions, the rules and techniques) instead of a long transcript. Nothing has to stay in context to be remembered.
• Helpers that keep the weight off. Retrievers read the long files and return one answer, coders run the edit, build and test loops, and each expert starts its task with a small, fresh context. What they read never lands in your session, and the statusline counts what that kept out.

	Vanilla	Project Architect
Average context per request	459k to 547k tokens	47k to 69k tokens
Context when the work starts	100k to 300k tokens	about 15k tokens
Cost per request, at API prices	$0.35 to $0.42	$0.08 to $0.11

Vanilla: real projects worked in one long Claude Code session. Project Architect: its own first phases, measured the same way. The whole story, with the measurements.

Learn more

The wiki has the rest: the statusline and the savings, the bench that measures the agents themselves (results in bench/results/), why PA3 stays serial, and the FAQ. What changed in each release: CHANGES.md.

License

See LICENSE.