Mitos
概览
作者提供的说明（英文原文）
Architectural decision memory for LLM-native workflows — markdown for humans, a graph for agents.
PopNx MCP 的三大支柱
这三项永远不会合并成一个分数：意见、测量和人气是不同的东西。
- 用户评价i暂无评价评价功能将在后续阶段与账户功能一起开放
- 我们的基础检查i1 项中通过 1 项依据公开信息检查；尚未扫描代码
- 使用情况i非常新↓ 88/周 · ★ –
它能帮你做什么
我们还没有为它写通俗的介绍。下面是我们从注册中心读取到的事实。
- 连接方式: 在你的电脑上运行
- 传输方式: stdio
- 分类: 搜索与 AI 记忆
安装
选择你使用的 AI 应用，然后复制设置。
安装之前
- 基础检查：1 项中通过 1 项
- 在你的电脑上运行，拥有你的用户账户的权限（文件、网络、应用）。
- 注册中心没有列出需要填写的密钥。
你需要
- PyPI已安装 Python 的 uv（用于 uvx 命令）
你用的是哪个应用？
在上方选择你的应用。我们会记住它，下次按钮会显示“安装到 …”。
打开 Claude Desktop → Settings → Developer → Edit Config，把下面的内容粘贴到 claude_desktop_config.json
{
"mcpServers": {
"mitos": {
"command": "uvx",
"args": [
"mitos-adr"
]
}
}
}在终端中运行这条命令
claude mcp add mitos -- uvx mitos-adr或者自己复制设置
粘贴到项目中的 .cursor/mcp.json，或 ~/.cursor/mcp.json
{
"mcpServers": {
"mitos": {
"command": "uvx",
"args": [
"mitos-adr"
]
}
}
}浏览器会先询问是否打开应用，然后应用会显示设置并请你确认。密钥的值会留空，由你自己填写。
或者自己复制设置
粘贴到项目中的 .vscode/mcp.json
{
"servers": {
"mitos": {
"type": "stdio",
"command": "uvx",
"args": [
"mitos-adr"
]
}
}
}浏览器会先询问是否打开应用，然后应用会显示设置并请你确认。密钥的值会留空，由你自己填写。
这个服务器在你的电脑上运行，而 ChatGPT 只能连接互联网上的服务器。请选择其他应用。
你的选择通过 Cookie 保存在这台设备上，无需账户。
这些设置是根据作者在注册中心的条目生成的。如果 README 的写法不同，请以 README 为准。
它能做什么
它在你的电脑上运行；安装后，工具列表会显示在你的 AI 应用中。
基础检查i
我们检查的是任何人都能看到的信息；我们没有阅读代码。不适用的检查项不计入总数。 · 检查于 2026年11月8日
- ✓ 通过
作者与仓库一致
注册名称已通过 GitHub 账户 dovahkiin-v 验证，该账户拥有此仓库
- … 尚未检查
仍在维护
我们还没有从 GitHub 读取这个仓库
- … 尚未检查
有许可证
我们还没有从 GitHub 读取这个仓库
- – 不适用
端点有响应
它在你的电脑上运行，没有互联网端点
- – 不适用
使用 OAuth 登录
在你的电脑上运行，改用本地设置
- … 尚未检查
工具描述中没有隐藏指令
它在你的电脑上运行；我们只能读取互联网端点的工具列表
与“搜索与 AI 记忆”中的其他服务器比较i
- 星标数多于此分类中其他 5716 个服务器的 0%
- 发布到注册中心的时间比此分类中其他 5716 个服务器的 67% 更新
- 基础检查结果等于或优于此分类中其他 5716 个服务器的 100%
与注册中心里这个分类中所有活跃的服务器比较。只有当发布者拥有该仓库时，星标数才计入。
常见问题
Mitos 安全吗？
它在 1 项适用的基础检查中通过了 1 项。我们没有阅读代码。在添加密钥或令牌之前，请先看看这个服务器要求哪些权限，并使用只具备所需权限的密钥。
它需要哪些密钥或代码？
注册中心没有列出必需的密钥，不过有些服务器会在首次使用时要求你登录。
它在我的电脑上运行还是在云端？
它在你的电脑上运行。你的 AI 应用会在需要时启动它，数据不会经过作者的服务器（除非它自己调用在线服务）。
它还在维护吗？
GitHub 仓库最后更新于–，最新版本（0.20.0）于20天前进入注册中心。
哪些应用可以使用它？
支持 MCP 的应用，例如 Claude Desktop、Claude Code、Cursor 和 VS Code。你的电脑上还需要有 Python (uv)。
有什么可以替代它？
同一分类中有：Context7, Codebase Memory, Exa
替代选择
同一分类中其他作者的服务器，按星标数排序。
如何看卡片：↓ 每周下载量i★ GitHub 星标i三大支柱：评价 · 检查 · 使用情况i更新 = 最新版本发布到注册中心的时间i
Codebase Memory
Codebase knowledge graph for AI agents — 162 languages, sub-ms queries, 99% fewer tokens.
Exa
Connect AI agents to Exa for web search, content fetching, and multi-step research.
作者的 README
打开 README
按原文从作者的仓库复制。PopNx MCP 没有审查或背书其内容。HTML、图片和脚本已被移除。链接会离开本站。
Mitos
🔧 Early release — actively developed
When you build software with AI assistants over months, the *reasoning* behind your decisions gets lost. The assistant forgets why you chose one approach, re-suggests options you already rejected, and your design notes drift out of sync with what was actually decided. Mitos is a memory layer for those decisions: it records each decision, the alternatives you ruled out, and how later decisions replace earlier ones — then feeds that history back to your AI assistant in a compact, trustworthy form.
The result: your AI collaborator stays consistent with the calls you've actually made — it stops contradicting a past decision or re-opening a settled question, and your decision record never silently rots.
Under the hood: markdown for humans (decisions.md and its quarterly archives under decisions/archive/ are the source of truth you can always read and grep), a typed graph for the agents (SQLite + a local Qdrant for semantic recall), and an MCP server so agents check precedent before deciding and record decisions as they make them.
Available on PyPI and the MCP Registry.
Fastest install: hand it to your agent
If you work with an AI coding agent (Claude Code, Cursor, Gemini CLI, …), the easiest path is to let it do the setup. In the project you want mitos in, give your agent:
Read https://github.com/dovahkiin-v/mitos/blob/main/SETUP.md and set up mitosfor this project. When done, run mitos status . from the project directory and report the result.
What your agent will end up doing — the same steps a human follows, all in SETUP.md where you can read them first:
- install the
mitosCLI via pipx (from PyPI or this repository); - start a local Qdrant container (
qdrant/qdranton port7333, isolated from any Qdrant you already run); - register the MCP server once for the whole machine, if it isn't registered already;
- initialize the project workspace, which also registers the project by name;
- ask you to set your API keys yourself (
mitos set-key) — a Gemini key (required), and an Anthropic key for the conflict-audit layer (strongly recommended); the setup guide tells agents not to handle key values.
How much your agent asks along the way is governed by your own agent's settings, not by this prompt.
Manual setup
The same steps by hand — full detail in [SETUP.md](SETUP.md):
- Install (once per machine):
pipx install mitos-adr - Start Qdrant (once per machine, shared by all projects):
docker compose up -dfrom this repo — mitos runs its own instance on:7333, so it never touches a Qdrant you use for other work. - Register the MCP server (once per machine, recommended for agents):
claude mcp add --scope user mitos -- mitos serve. One registration serves every project — see SETUP.md for what it costs, for other harnesses, and for why a leftover per-project.mcp.jsonentry has to go. - Per project:
mitos initfrom the project root, thenmitos set-key --global(one key covers everything; get it at https://aistudio.google.com/app/apikey). Gemini is the tested embedding provider today; a multi-provider abstraction is on the roadmap. - Verify:
mitos status .→READY ✓.
mitos status . is the compass throughout: it says exactly what's done, what's missing, and what to do next for that project. With no project named, mitos status answers the other question — what does this machine have — listing every registered project and checking Qdrant.
Every command names its project. There is no default target: mitos init registers the project by name, and from then on each verb takes -p , -p , or -p . from the project root (agents pass the same thing as a project argument). mitos projects lists what's registered, and mitos unregister removes a registration (the workspace stays as it is). That is what lets one install and one MCP server serve every project on the machine without a call ever landing in the wrong corpus.
How it runs
Mitos is per-project — each project gets its own decision graph and its own Qdrant collections. Day to day, three verbs carry the loop (as MCP tools for agents, with CLI twins):
| Verb | When |
|---|---|
surface_decisions (mitos surface) | *Before* deciding — is there precedent? Every hit carries the alternatives that were already rejected and why. |
record_decision (mitos record) | The moment something is settled — the decision, the rejected paths, and how it relates to prior decisions (supersedes, amends, …). |
query_decisions (mitos query) | Looking something up by meaning (the MCP tool also by exact handle), or searching only the rejected options. |
A few properties worth knowing:
- The markdown is the source of truth. Every decision lands in
decisions.md, and rotation moves committed entries intodecisions/archive/; both are human-readable and greppable, and the graph and the search index are derived from them and can always be rebuilt from them (mitos rebuild). - A decision's axiom is never edited — it's superseded. The axiom and its mechanisms change only through a later decision that supersedes, corrects or amends it, and state (active / superseded / amended) is computed from those typed relations, so the history of *why* always survives. Commentary (rejected paths, context, scope) can be fixed in place with
mitos amend-commentary. - It fails safe. If the search index or the embedding API is down, recording still works and search degrades to an honest text-match over the markdown — nothing blocks, nothing is lost, and degraded output says it's degraded.
- It audits itself. The corpus sweep (
mitos check -p .) finds decisions that silently contradict each other. The commit gate (mitos hook-install -p .) runs ongit commitand holds a commit made after decisions were recorded until a check has been attempted, and--stagedchecks hand-authored entries still pending indecisions.md— see SETUP.md for the commit gate, the staged gate, and the CI and cron recipes.
Explore the rest with mitos --help — the help text doubles as the API reference.
Why it exists
Building software through intensive LLM design reviews produces architectural decisions faster than a person can track. One month of that working style produced close to 900 decision records in a single markdown file — no longer greppable, readable, or manageable by hand. Existing ADR tools are built for human teams logging the occasional decision; mitos is built for a solo developer whose AI assistants generate and consume decisions continuously.
If that's your way of working, project size doesn't matter much — the higher the decision volume, the faster mitos moves from comfort to necessity.
Development
pip install -e '.[test]'MITOS_NO_LIVE_TESTS=1 pytest -m "not packaging" -n auto # offline suite, parallel (~50s) pytest -m "not packaging" # adds the live tier — serial only pytest -m packaging # real-install check: fresh venv + pip install
-n auto is safe for the offline suite and not for the live tier: the test-collection sweep is session-scoped, so parallel workers delete each other's Qdrant collections and the affected tests degrade to skips rather than failures.
The *_live.py suites and golden Layer B make real Gemini and Anthropic API calls against your own keys, and need Qdrant on :7333. They skip when no key is resolvable, so a fresh clone runs the fast path by default.
Keys resolve from the environment, a repo-root .env, or ~/.config/mitos/.env — so if you already *use* mitos, a test run can pick up your personal key and spend against it. Opt out explicitly:
MITOS_NO_LIVE_TESTS=1 PYTHONPATH=. pytest -m "not packaging"The canonical decision format lives in mitos/format-spec.md. License: Apache 2.0.