Plan
БесплатноНе проверенA control plane for trustworthy AI agents: write product judgment once, then enforce it through typed MCP tools, git gates, demolition-first stories, evidence-j
Описание
A control plane for trustworthy AI agents: write product judgment once, then enforce it through typed MCP tools, git gates, demolition-first stories, evidence-judged checkboxes, and zero-trust monitoring. Ships with Claude Code and Cursor plugins, Agent Skills, and a FastMCP server.
README
Kyzo Plan
You decide what's worth building and write it once — agents execute against that
judgment, gated on git, with a judge before every checkbox and a token meter kept lean by design.
Kyzo Plan is a control plane for people who still own the product judgment — and for agents that must not re-derive it. Intent is written once, upstream, as an execution contract. The board is shared state, not a handoff chain. 35 typed MCP tools operate it; git and a completion judge keep motion honest. A checkbox is a fact, not a self-attestation.
Install
Requirements: uv on PATH, gh authenticated against
the board's GitHub org, Python 3.12+.
Claude Code
/plugin marketplace add https://github.com/kyzodb/plan
/plugin install kyzo-plan@kyzo
Zero config in the common case: board owner and repo come from origin, project
number from the repo's sole open linked GitHub Project. create_board provisions
a fresh board with this schema if you're starting from nothing.
Overrides on enable (/plugin configure kyzo-plan@kyzo) or at install:
claude plugin marketplace add https://github.com/kyzodb/plan
claude plugin install kyzo-plan@kyzo \
--config board_owner=OWNER --config board_repo=REPO --config board_project=N
Cursor — same skills, agents, and MCP server. Claude Code is the supported install from this repository today.
Uninstall with /plugin uninstall kyzo-plan@kyzo. Board state lives on GitHub;
uninstalling leaves nothing behind.
One ledger — no translation layer
What you see is what agents operate. Stock GitHub Projects; one screen is the working truth — not a status chain between roles:

Epics carry roll-up progress (1 / 3 — 33%). Classification is the GitHub label.
One card in In Progress answers what is being built right now. Filter Backlog
and Done away:

Where the tokens go
Agentic cost is mostly input: context re-read on every step. Five drains, five mechanisms — thinking stays upstream; agents don't burn the meter rediscovering it:
One axis, horizons on epics
A grammar, not a manual
The kyzo-plan-manage-board skill carries what lives between the tools: orient
with read_board_status, never work around a refusal, never touch the board
through raw gh.
The lifecycle refuses until reality agrees
Branch-per-epic, one story at a time. The board cannot outrun the repository.
start_epic demands a clean, up-to-date main and an unused branch. start_story
demands HEAD on that branch, stories in sub-issue order, and prior work
physically present. finish_epic refuses until every box is checked and the
branch has no commit missing from main — it verifies the merge; it never
performs it.
Judgment in, once
The hard work is still yours: what this serves, what dies, what was chosen, and
how done is proven. Skills kyzo-plan-write-story and kyzo-plan-write-epic
hold the shapes. You commit:
- Sources / Condemned / Ceiling / Engineering Choice — the product call, crystallized so an agent cannot reopen it as exploration.
- Context — exact references; only genuine unknowns marked
[OPEN]. - Tasks — append-only
T#withAllowlist:and a fastCheck:. - Definition of Done — exactly one Final QA item (
check_final_qa). No Witness/CI/worktrees in the contract. Allowlist commit is the seal.
Banned lexicon (improve, polish, for now…) stays out of tasks — or lives only inside Condemned, naming what is being killed. Agents execute the contract; they do not re-author the judgment.
Watched at the call
The spawner runs under zero trust: this agent will deviate; an unwatched deviation is the spawner's fault. Paths are law. A live session:

Orchestration lives in kyzo-plan-run-story: allowlist arming, XML spawn, path
watch, parent check, judge after verify_task_completion, allowlist commit (the seal),
Final QA + check_final_qa.
Built by Kyzo
Kyzo Plan is the system KyzoDB is built with — these screenshots are its live board. It does not use or require KyzoDB. We needed a control plane we could trust with agents; we built it, we run it every day, and it earned a release of its own.
Next in the line: Codegraph — measuring whether each change moves a codebase toward the architecture its team intends.
License
Business Source License 1.1 — free to use, modify, and build on for any non-production purpose; production use requires a commercial license until the Change Date, after which it converts to MPL-2.0. See LICENSE.
Установка Plan
У этого сервера нет опубликованного пакета — он собирается из исходников. Открой репозиторий и следуй инструкции в README.
▸ github.com/kyzodb/planFAQ
Plan MCP бесплатный?
Да, Plan MCP бесплатный — установка в пару кликов через Unyly без оплаты.
Нужен ли API-ключ для Plan?
Нет, Plan работает без API-ключей и переменных окружения.
Plan — hosted или self-hosted?
Self-hosted: сервер запускается локально на твоей машине командой из раздела установки.
Как установить Plan в Claude Desktop, Claude Code или Cursor?
Открой Plan на unyly.org, выбери вкладку своего клиента (Claude Desktop, Claude Code, Cursor) и нажми Install — конфиг сгенерируется автоматически, без правки JSON.
Похожие MCP
GitHub
PRs, issues, code search, CI status
автор: GitHubFilesystem
Secure file operations with configurable access controls.
Memory
Knowledge graph-based persistent memory system.
Template MCP Server
A CLI tool to create a new Model Context Protocol server project with TypeScript support, dual transport options, and an extensible structure
автор: mcpdotdirectCompare Plan with
Не уверен что выбрать?
Найди свой стек за 60 секунд
Автор?
Embed-бейдж для README
Похожее
Все в категории development
