Plans, red-teams and checks your project.
Marvora Blueprint is a free plugin for Claude Code, under the MIT licence. It takes your project through a staged method with gates where you decide, and hands you a plan. It does not build.
Install: paste both lines into your terminal
claude plugin marketplace add Rocky-Ventures/marvora
claude plugin install marvora@marvora
Then run /marvora:blueprint.
Published by Rocky Ventures from its GitHub organisation, Rocky-Ventures: the only official source. No hooks, MCP servers or installers; github.com/marvora and any npm package named marvora are not ours.
Release commit: 4b74dfa
Before you install
- A paid Claude plan or API billing. The agents run on your plan's usage, or on your money if you pay per token. Free plans have no Workflow tool.
- The Workflow tool, for standard depth. It launches a team of agents. On Pro, turn on the "Dynamic workflows" row in
/configfirst. Without it you get quick and light depth only. - macOS, Linux or WSL2. WSL2 is Linux inside Windows. Only macOS has been run live by us. Native Windows is not supported.
shandpython3. The preflight and forget scripts use them.
One method, three gates
Scouts read your project and the ground around it. Discovery turns what they find into a brief. Designs are scored by judges and attacked by a red team, errata are fixed, and the plan is finalised.
At each gate Blueprint stops and asks for your decision. What you get is a plan, written for people or for your own agents to build from.
See a run before you start one
Two real transcripts from a small invented project, run headless with scripted asks and no person in the loop: one at quick depth, one at light depth. Standard depth, with its judges and red team, has not yet been run from an install.
-
Quick depth
About 1 minute, no agents: the biggest risk, the one fix to make first, and what was not checked.
-
Light depth
About 2 minutes, no agents launched: what was found, four options with their effort and catch, and one question for you.
On your machine
- It reads anything you can read.
- It spends your plan's usage, or your money if you pay per token. There is no spend cap in code: on metered billing, set a limit in your provider console and start with light depth.
- The default fence lets every agent fetch the listed domains unasked, so an injected page there could send out what an agent read.
- It leaves files on your disk.
/marvora:blueprint forgetlists what it can find and deletes nothing. The Remove everything steps in the README delete them by hand, with or without the plugin. - It sends no telemetry. Research questions go to web search and to the sites it fetches, and the fence probe tries one request to example.com in each form, which gets out only if the fence is not working.
- Light depth runs on your current permission rules, with no sandbox and no fence.
- Role rules are guidance to the agents, not limits enforced in code.
- A project's CLAUDE.md and settings reach every agent, so use it only in a project you trust.
Proof
Each depth
- Quick depth
- Run end to end by us
- Light depth
- Run end to end by us
- Standard depth
- Never run from an install: you would be first
- Full depth
- Never run from an install: you would be first
- Run end to end by us
- Run by us on macOS
- Never run from an install: you would be first
Row by row
| What | Status | Detail |
|---|---|---|
| Install by the two commands | Run end to end by us | From a local folder holding the release candidate, not yet from GitHub; claude plugin validate . --strict passes |
| Quick depth | Run end to end by us | One run, about 1 minute, 0 agents; transcript on the sample page |
| Light depth | Run end to end by us | One run, about 2 minutes, 0 agents launched (the project was one small file, so the orchestrator read it itself); transcript on the sample page |
| Preflight script | Run by us on macOS | Its tests also pass in a Linux (Alpine) container |
| Forget script | Run by us on macOS | Nothing to list, since quick and light runs make no run folder; its tests also pass in a Linux (Alpine) container |
| Standard depth: discover, design, errata, late input, finalise | Never run from an install: you would be first | Covered only by offline contract tests (a mock harness) |
| The fence card, the sandbox probe, any workflow launch | Never run from an install: you would be first | Offline contract tests only |
| Full depth | Never run from an install: you would be first | Off unless your own config allows it |
| Linux and WSL2, live | Never run | Only the preflight and forget script tests ran in a container |
| Native Windows | Not supported | |
| Pro, Max and API billing | Never run by us | The one account used shows an Enterprise plan |
No agent count, minute count or prompt count exists for standard or full depth from this plugin. No dollar figure was measured. The time and cost figures in the plugin's depth reference (depth.md) and wherever they are repeated come from unpublished runs of the method made by its author before this plugin existed, on another setup; none comes from Marvora Blueprint 0.1.0 or from any install of it.
Conditions of the run
macOS, Claude Code 2.1.327, one machine. A fresh, empty Claude Code configuration folder, signed in to the publisher's own account (plan tier shown by the client: Enterprise). The plugin was installed by the two commands from a local folder holding the release candidate, not yet from GitHub. The runs were headless (claude -p), with scripted asks and no person in the loop. Headless mode cannot answer permission prompts, so only Read, Glob and Grep were allowed (plus the Agent tool and sh for the light run), and no fence card, sandbox probe or workflow launch took place. The project was a small invented booking app, Tidy Table: one 10-line Node server and a README, with a planted fake credential file, which did not appear in either transcript. These numbers come from one run each on one setup, not yours.
What is coming
Announced, with no date or price.
- Build is next, in 0.2.0. Its bar is the publisher's alone.
- Tool limits by role. Each agent gets only the tools its role needs.
- A paid tier. Version 0.1.0 stays free under MIT.