kumihimo
Enables agents to read, restructure, validate, and braid plan graphs of plain-text files, allowing them to maintain and execute the plan interactively.
README
Kumihimo
組紐 — many threads in, one cord out.
Kumihimo is the Japanese craft of braiding many separate threads into a single cord. This tool works the same way: you lay out a plan as a graph of plain-text files, and Kumihimo braids that graph into a single, well-ordered prompt an LLM or coding agent can act on.
<img src="docs/assets/canvas-editor.png" alt="The Kumihimo editor: a plan graph with kind-colored nodes, dependency and membership edges, and the selected node's form in the sidebar" width="100%">
-
See it.
kumihimo editopens a live canvas — kind-colored nodes, dependency/membership/annotation edges drawn distinctly, drag to arrange, click to edit. The canvas follows the files: change a node in vim (or let Claude change it over MCP) and watch it move. -
Braid it.
kumihimo braidcompiles the graph into one deterministic prompt: topologically ordered, sectioned by milestone, every item stating its real prerequisites, the graph's own Mermaid diagram embedded. -
Let Claude drive it.
kumihimo mcpserves the plan over MCP — eleven tools to read, restructure, validate, and braid — so an agent can maintain the plan it executes. This repo's own roadmap is a Kumihimo plan (plans/roadmap), its milestone prompts were braided, not hand-written, and the canvas below is that roadmap the morning v0.1 finished — every thread done, the release node blocked on a human with a tag:<img src="docs/assets/canvas-roadmap.png" alt="Kumihimo's own v0.1 roadmap on the canvas: milestones M3-M6 with every task done and the release node blocked" width="100%">
The files are the only source of truth — Markdown with YAML frontmatter, in
git, diffing cleanly. The editor, CLI, and MCP server are thin clients over
one operations layer; node positions live in a view.yaml sidecar so
arranging the graph never dirties a semantic diff.
Ten minutes
Kumihimo is not on a package index yet — install from source (Python 3.11+; Node 18+ only if you want the canvas):
git clone https://github.com/tzavislan/kumihimo
cd kumihimo
cd frontend && npm install && npm run build && cd ..
pip install .
The npm line builds the live canvas and can be skipped — everything below
still works except kumihimo edit, which then serves the API with an honest
"frontend not built" page instead of the graph.
kumihimo new myplan
kumihimo edit myplan # live canvas at 127.0.0.1:8720
kumihimo add myplan build --kind task --title "Build the loader" --body "..."
kumihimo check myplan # cycles named with paths, dangling edges, field breaches
kumihimo braid myplan # one prompt on stdout — paste into your agent
A node is a file:
---
kind: task
title: Rate-limit middleware
needs: [api-endpoints, pick-algorithm]
in: [ship-guarded-api]
effort: M
acceptance:
- 429 + Retry-After on breach
---
Middleware on every authenticated route. Fail *open* on Redis errors —
availability beats enforcement.
…and the braid compiles the graph into one prompt — preamble, embedded Mermaid shape, settled decisions stated as constraints, every task with an honest After: line and Done when: checklist:
<img src="docs/assets/braid-modal.png" alt="The braid preview modal: the compiled prompt with its embedded Mermaid diagram and reading rubric" width="100%">
The core is domain-agnostic: it understands identity, prose, order (needs),
membership (in), and annotation (links) — everything else comes from
kinds you define in the manifest, fields and Jinja2 templates included.
An engineering pack (task, milestone, decision, risk, question) ships as the
default; examples/fieldnotes is a research plan whose
kinds live entirely in its own manifest.
Docs: https://tzavislan.github.io/kumihimo/ — tutorial, concepts, custom kinds, MCP setup, file formats.
Development
uv sync
uv run pytest -q
uv run python tools/lint.py # the conventions linter CI runs
cd frontend && npm install && npm run build # only if you work on the canvas
Conventions (enforced in CI): CONVENTIONS.md. Contributions: CONTRIBUTING.md. Design record: PLAN.md. Releases: RELEASING.md.
License
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。