AgentPlugged

AgentPlugged is a method and an engine for running a team of AI agents on your product, from the idea to the release. You talk to one assistant, Noah, in plain language; he hands each request to the agent that owns it (brief, PRD, architecture, stories, code, reviews) and brings you each decision one at a time, with a recommendation.

What sets it apart: the rigor lives in the engine, not in long prompts. Every decision, checkpoint and result is an entry in an append-only journal inside your Git repository, so an agent cannot skip a step, and every story ends with a proof you can read. It also keeps the token cost low, because each agent receives only the passages it needs.

It is free, needs no account and no server: npx agentplugged init in a project, then work from Claude Code.

Your first ten minutes

You need Node.js 24 or later and Claude Code. A Git repository is recommended (the journal lives on a branch of it). Open a terminal in a new folder, or in a project you already have.

1. Install the method (about 2 minutes).

npx agentplugged init

init first looks at your machine (Node, Git, your AI tools, the keychain) and offers to repair what is missing; it shows the exact command and runs it only after one key. Then it asks a few questions in keyboard menus (arrows, space, Enter; Escape goes back). Every question has a recommended answer: pressing Enter all the way gives a correct installation. The profile question is the one that matters:

• Complete new product: you start from an idea, with framing and planning;
• Existing product in maintenance: your code is already there, framing is skipped and the assistant first writes a knowledge base of your project.

What you see: the list of what was installed in .agentplugged/, your choices, then the last line, which names your tool: Ready. Open Claude Code in this folder and write: noah ?. In a script, npx agentplugged init --yes asks nothing and takes the recommended answers. Nothing is ever written over your AGENTS.md or CLAUDE.md: a short pointer is added once.

2. Say hello to Noah (1 minute). Open Claude Code in the same folder and write:

noah ?

What you see: a message starting with 🧭 Noah :. He reads the state of the work, tells you in a sentence where you stand (a new product: nothing yet) and hands over to the agent that comes next. You never repeat what is already written down.

3. Frame your idea (about 5 minutes). For a new product, Iris, the product analyst, takes the conversation and asks you questions one at a time, then writes the brief (.agentplugged/docs/brief.md). Answer in your own words. When the brief is ready, Noah presents it with his recommendation and a closed question: answer with a word ("ok", "yes", or what you want changed). That answer is your first validation, recorded in the journal.

4. See where you stand, at any time.

npx agentplugged next

What you see, in three lines: the current step, the agent to call and how (/iris in Claude Code), and what waits for you. When a step has a command, a fourth line gives it.

5. Your first story, and its proof. After the brief come the PRD (Maya), the architecture (Alex) and the breakdown into small stories (Izzy); you validate each document in the chat. Then a developer builds a story in its own Git branch and working copy, and a reviewer checks it. A story is done when it comes back with its proof:

npx agentplugged trace     # each requirement with its status, down to the stories and tests that cover it
npx agentplugged history   # your validations: what, when, with the proof of each

If something does not work, say what you saw to Noah in the chat, or run npx agentplugged doctor: it names what is missing on your machine and the command that repairs it.

Essential commands

Command	What it does
npx agentplugged init	Installs or refreshes the method; your documents, stories and settings are kept.
npx agentplugged sync	Writes the files your AI tool reads (skills, agents, AGENTS.md). init does it for you.
npx agentplugged next	The next step, the agent to call, what waits for you.
npx agentplugged doctor	Checks your machine, repairs what is missing with your agreement. --report prints a report to attach to an issue, with nothing personal in it.
npx agentplugged trace	Requirements and their status, down to the proof.
npx agentplugged history	Your validations, most recent first.
npx agentplugged --help	Every command and option.

Full documentation of the commands: apps/cli/README.md. The team of agents: method/README.md.

Honest limits

• Claude Code comes first. It is the tool where the method is complete. Codex, Antigravity CLI, Cursor and Kimi Code CLI are supported by sync (their files, their MCP configuration), but with fewer guards: where a tool cannot enforce a rule, the rule is written in the constitution and sync says so. They improve little by little.
• You need Node.js 24 or later. On Linux, the keys live in the Secret Service (GNOME Keyring, KWallet or KeePassXC), which must be running and unlocked.
• The agents run in your own AI tool, on your own subscription: AgentPlugged adds no model of its own.
• A permission rule that an agent deliberately works around is out of reach of a check; the strict validation level exists for that.

What is in this repository

Folder	Content
apps/cli	The agentplugged command (npm package agentplugged).
apps/mcp	The MCP server the agents report their status, progress and results through.
packages/	The libraries of the engine: core, protocol, crypto, i18n, adapters-git, adapters-tools, adapters-syntax.
method/	The agents, skills and models.yaml that init installs.
tooling/	The checks on the code: import boundaries, the test of the published package, and the install trials.

Build and test

This is a pnpm workspace (pnpm 12.8.1, pinned by packageManager; Node.js 24 or later).

pnpm install --frozen-lockfile
pnpm typecheck
pnpm lint
pnpm build
pnpm exec turbo run test --concurrency=2
pnpm smoke     # packs the CLI, installs it in a new folder and runs it

The installation is also tried on every change and on the published version, on Windows, macOS and Linux, in the usual terminals and on imperfect machines (no connection, no Git identity, a long path with accents, a proxy): see the Install trials workflow, or run pnpm trials on your machine.

License

AgentPlugged has two licenses, one for each part:

• The engine (apps/cli, apps/mcp, packages/, tooling/) is under the Functional Source License, Version 1.1, MIT Future License (FSL-1.1-MIT). You can use it, read it and modify it, but not use it for a competing product or service. Each version becomes MIT two years after its publication.
• The method (method/) is under the MIT license.

LICENSE at the root of the repository explains the split. Copyright 2026 Innovel Studio SASU, publisher of AgentPlugged.

Contributing

Issues and pull requests are welcome on GitHub. Before opening a pull request, run the commands of "Build and test" above: the CI runs the same ones on Linux and Windows. For a larger change, open an issue first so that it can be discussed.