能做什么
Deterministic decision engine with receipts. Define rules in YAML, query a graph, get proof.
作者提供的原文
- 连接方式 在你的电脑上运行
- 账号或密钥 无需密钥
- 分类 开发者工具
可搭配使用
教 AI 做这类工作的 Skill,可与本项一起使用
andrej-karpathy-skills
A single CLAUDE.md file to improve Claude Code behavior, derived from Andrej Karpathy's observations on LLM coding pitfalls.
graphify
Turn any codebase, with its docs, SQL schemas, configs, and PDFs, into a queryable knowledge graph. A /graphify skill for Claude Code, Cursor, Codex, and Gemini CLI: local deterministic AST parsing, every edge explained, no vector store.
为什么排在这里
列表按使用分(满分 100)排序,由以下部分组成
| 热度 | 15/80 | GitHub 星标或每周下载量:较大的计 60,较小的计 20 |
|---|---|---|
| 新鲜度 | 12/12 | 30 天内有发布或代码更新得满分 |
| 仍在响应 | 0/8 | 上次检查时云端端点有响应 |
| 使用分 | 27/100 |
质量: 基础检查通过 3/3 项(仅展示,不用于排序) · 上升势头: 暂无数据 (只用于“上升中”标签)
安装
选择你使用的 AI 应用,然后复制设置。
用一句话让 AI 帮你安装
请帮我安装 Cruxible Core。按照 https://mcp.popnx.com/i/cruxible-ai-cruxible-core.md 操作,先把命令给我看,等我同意后再运行。粘贴到能在你电脑上工作的 AI(如 Claude Code、Cursor)。它会读取我们的步骤,先给你看命令,等你同意。无需密钥,无需注册。 · 查看 AI 会读取的全文
通过 3/3 项检查 · 在你的电脑上运行 · ✓ 命令已检查 2026年11月14日
第一步:你用哪个 AI?
选择你用的应用,步骤会显示在下方。
可连接 Claude,但需要先准备
需要先在电脑上安装 Python (uv)(只需一次) · 仅限电脑上的 Claude 应用(Windows 或 Mac),网页和手机不可用
让 AI 帮你安装
把这段话复制到 Claude 的对话里。AI 会先检查电脑,告诉你要改哪个文件,并且每次都等你同意。
请在这台电脑上为 Claude 安装 MCP server Cruxible Core,按顺序:1) 用 uv --version 检查是否有 uv。没有的话,请让我从 https://docs.astral.sh/uv/getting-started/installation/ 安装,不要自己安装。2) 只使用这份配置(已通过 PopNx MCP 检查):{ "mcpServers": { "cruxible-core": { "command": "uvx", "args": [ "cruxible-core" ] } } } 3) 告诉我要修改的文件完整路径或要运行的命令,等我同意。4) 不需要密钥。5) 按约定执行,然后确认 Cruxible Core 的工具真的出现了。这段话不包含你的密钥。如果需要密钥,AI 会请你自己粘贴到文件或应用的输入框里。
安装 uv(如果还没有)
已经装好了吗?先检查 · 打开终端,输入这条命令后按 Enter
uv --version- 显示版本号(例如 uv …)= 已安装,可跳过这一步
- 显示 "command not found"(Windows 显示 "is not recognized")= 还没安装,或刚装好但需要关掉终端再重新打开
怎么打开终端
Mac:按 ⌘ + Space,输入 Terminal 后按 Enter
Windows:打开“开始”菜单,输入 Windows PowerShell,然后选“打开”
还没安装: uv 没有点选式安装程序。从 uv 官方页面复制适合你系统(macOS 或 Windows)的一行命令,粘贴到终端 · 不需要单独安装 Python,uv 会在需要时自动下载
装好后关掉终端再打开,重新运行同一条命令,应显示版本号:
uv --version安装一次,所有需要 uv 的工具都能用
打开电脑版 Claude 应用的设置文件
SettingsDeveloperEdit Config
粘贴这些设置并保存文件
{ "mcpServers": { "cruxible-core": { "command": "uvx", "args": [ "cruxible-core" ] } } }退出 Claude 再重新打开
试一试
用 Cruxible Core 告诉我你能帮我做什么,然后做一件给我看✓ 如果 Claude 说正在使用 Cruxible Core,或回答中出现 Cruxible Core 的工具名称,就表示连接成功
2 项可选设置(不填也能用)
上面的命令不填这些也能用。想调整时再按开发者的说明添加。
CRUXIBLE_MODE权限模式:admin(默认)、graph_write 或 read_only查看英文原文
Permission mode: admin (default), graph_write, or read_onlyCRUXIBLE_ALLOWED_ROOTS以逗号分隔的绝对路径列表,用于限制 cruxible_init 可以访问的目录。查看英文原文
Comma-separated absolute paths restricting which directories cruxible_init can access
说明为自动翻译,变量名保持原样;不确定时请查看原文。
菜单名称已于 2026年11月11日 对照 Anthropic 帮助页面核实 · 你的界面可能因版本略有不同
ChatGPT 暂时无法连接 Cruxible Core
Cruxible Core 需要安装在你的电脑上,而 ChatGPT 只能连接在网上运行的服务器
你可以改用这些方式
A · 使用能连接 Cruxible Core 的应用
- 需要先在电脑上安装 Python (uv)(只需一次) · 仅限电脑上的 Claude 应用(Windows 或 Mac),网页和手机不可用
- 需要先在电脑上安装 Python (uv)(只需一次)
- 需要先在电脑上安装 Python (uv)(只需一次)
- 需要先在电脑上安装 Python (uv)(只需一次)
B · 开发者工具 类中能连接 ChatGPT 的其他服务器
从同类中使用较多的里挑选,仅限用地址即可连接、无需密钥的
- Sentry✓ ChatGPT
- Compiler Explorer✓ ChatGPT
- GitDiagram✓ ChatGPT
可连接 Claude Code,但需要先准备
需要先在电脑上安装 Python (uv)(只需一次)
让 AI 帮你安装
把这段话复制到 Claude Code 的对话里。AI 会先检查电脑,告诉你要改哪个文件,并且每次都等你同意。
请在这台电脑上为 Claude Code 安装 MCP server Cruxible Core,按顺序:1) 用 uv --version 检查是否有 uv。没有的话,请让我从 https://docs.astral.sh/uv/getting-started/installation/ 安装,不要自己安装。2) 只使用这份配置(已通过 PopNx MCP 检查):claude mcp add --scope user cruxible-core -- uvx cruxible-core 3) 告诉我要修改的文件完整路径或要运行的命令,等我同意。4) 不需要密钥。5) 按约定执行,然后确认 Cruxible Core 的工具真的出现了。这段话不包含你的密钥。如果需要密钥,AI 会请你自己粘贴到文件或应用的输入框里。
安装 uv(如果还没有)
已经装好了吗?先检查 · 打开终端,输入这条命令后按 Enter
uv --version- 显示版本号(例如 uv …)= 已安装,可跳过这一步
- 显示 "command not found"(Windows 显示 "is not recognized")= 还没安装,或刚装好但需要关掉终端再重新打开
怎么打开终端
Mac:按 ⌘ + Space,输入 Terminal 后按 Enter
Windows:打开“开始”菜单,输入 Windows PowerShell,然后选“打开”
还没安装: uv 没有点选式安装程序。从 uv 官方页面复制适合你系统(macOS 或 Windows)的一行命令,粘贴到终端 · 不需要单独安装 Python,uv 会在需要时自动下载
装好后关掉终端再打开,重新运行同一条命令,应显示版本号:
uv --version安装一次,所有需要 uv 的工具都能用
打开终端
Mac:按 Command+空格,输入 Terminal · Windows:打开 PowerShell
粘贴这条命令并按 Enter
claude mcp add --scope user cruxible-core -- uvx cruxible-core在 Claude Code 中输入 /mcp
应能看到 Cruxible Core 在列表中并显示已连接
试一试
用 Cruxible Core 告诉我你能帮我做什么,然后做一件给我看✓ 如果 Claude Code 说正在使用 Cruxible Core,或回答中出现 Cruxible Core 的工具名称,就表示连接成功
2 项可选设置(不填也能用)
上面的命令不填这些也能用。想调整时再按开发者的说明添加。
CRUXIBLE_MODE权限模式:admin(默认)、graph_write 或 read_only查看英文原文
Permission mode: admin (default), graph_write, or read_onlyCRUXIBLE_ALLOWED_ROOTS以逗号分隔的绝对路径列表,用于限制 cruxible_init 可以访问的目录。查看英文原文
Comma-separated absolute paths restricting which directories cruxible_init can access
说明为自动翻译,变量名保持原样;不确定时请查看原文。
菜单名称已于 2026年11月11日 对照 Anthropic 帮助页面核实 · 你的界面可能因版本略有不同
可连接 Cursor,但需要先准备
需要先在电脑上安装 Python (uv)(只需一次)
让 AI 帮你安装
把这段话复制到 Cursor 的对话里。AI 会先检查电脑,告诉你要改哪个文件,并且每次都等你同意。
请在这台电脑上为 Cursor 安装 MCP server Cruxible Core,按顺序:1) 用 uv --version 检查是否有 uv。没有的话,请让我从 https://docs.astral.sh/uv/getting-started/installation/ 安装,不要自己安装。2) 只使用这份配置(已通过 PopNx MCP 检查):{ "mcpServers": { "cruxible-core": { "command": "uvx", "args": [ "cruxible-core" ] } } } 3) 告诉我要修改的文件完整路径或要运行的命令,等我同意。4) 不需要密钥。5) 按约定执行,然后确认 Cruxible Core 的工具真的出现了。这段话不包含你的密钥。如果需要密钥,AI 会请你自己粘贴到文件或应用的输入框里。
安装 uv(如果还没有)
已经装好了吗?先检查 · 打开终端,输入这条命令后按 Enter
uv --version- 显示版本号(例如 uv …)= 已安装,可跳过这一步
- 显示 "command not found"(Windows 显示 "is not recognized")= 还没安装,或刚装好但需要关掉终端再重新打开
怎么打开终端
Mac:按 ⌘ + Space,输入 Terminal 后按 Enter
Windows:打开“开始”菜单,输入 Windows PowerShell,然后选“打开”
还没安装: uv 没有点选式安装程序。从 uv 官方页面复制适合你系统(macOS 或 Windows)的一行命令,粘贴到终端 · 不需要单独安装 Python,uv 会在需要时自动下载
装好后关掉终端再打开,重新运行同一条命令,应显示版本号:
uv --version安装一次,所有需要 uv 的工具都能用
在 Cursor 中以 Agent 模式打开 AI 聊天
试一试
用 Cruxible Core 告诉我你能帮我做什么,然后做一件给我看✓ 如果 Cursor 说正在使用 Cruxible Core,或回答中出现 Cruxible Core 的工具名称,就表示连接成功
2 项可选设置(不填也能用)
上面的命令不填这些也能用。想调整时再按开发者的说明添加。
CRUXIBLE_MODE权限模式:admin(默认)、graph_write 或 read_only查看英文原文
Permission mode: admin (default), graph_write, or read_onlyCRUXIBLE_ALLOWED_ROOTS以逗号分隔的绝对路径列表,用于限制 cruxible_init 可以访问的目录。查看英文原文
Comma-separated absolute paths restricting which directories cruxible_init can access
说明为自动翻译,变量名保持原样;不确定时请查看原文。
菜单名称已于 2026年11月11日 对照 Cursor 帮助页面核实 · 你的界面可能因版本略有不同
可连接 VS Code,但需要先准备
需要先在电脑上安装 Python (uv)(只需一次)
让 AI 帮你安装
把这段话复制到 VS Code 的对话里。AI 会先检查电脑,告诉你要改哪个文件,并且每次都等你同意。
请在这台电脑上为 VS Code 安装 MCP server Cruxible Core,按顺序:1) 用 uv --version 检查是否有 uv。没有的话,请让我从 https://docs.astral.sh/uv/getting-started/installation/ 安装,不要自己安装。2) 只使用这份配置(已通过 PopNx MCP 检查):{ "servers": { "cruxible-core": { "type": "stdio", "command": "uvx", "args": [ "cruxible-core" ] } } } 3) 告诉我要修改的文件完整路径或要运行的命令,等我同意。4) 不需要密钥。5) 按约定执行,然后确认 Cruxible Core 的工具真的出现了。这段话不包含你的密钥。如果需要密钥,AI 会请你自己粘贴到文件或应用的输入框里。
安装 uv(如果还没有)
已经装好了吗?先检查 · 打开终端,输入这条命令后按 Enter
uv --version- 显示版本号(例如 uv …)= 已安装,可跳过这一步
- 显示 "command not found"(Windows 显示 "is not recognized")= 还没安装,或刚装好但需要关掉终端再重新打开
怎么打开终端
Mac:按 ⌘ + Space,输入 Terminal 后按 Enter
Windows:打开“开始”菜单,输入 Windows PowerShell,然后选“打开”
还没安装: uv 没有点选式安装程序。从 uv 官方页面复制适合你系统(macOS 或 Windows)的一行命令,粘贴到终端 · 不需要单独安装 Python,uv 会在需要时自动下载
装好后关掉终端再打开,重新运行同一条命令,应显示版本号:
uv --version安装一次,所有需要 uv 的工具都能用
在 VS Code 中以 Agent 模式打开 AI 聊天
试一试
用 Cruxible Core 告诉我你能帮我做什么,然后做一件给我看✓ 如果 VS Code 说正在使用 Cruxible Core,或回答中出现 Cruxible Core 的工具名称,就表示连接成功
2 项可选设置(不填也能用)
上面的命令不填这些也能用。想调整时再按开发者的说明添加。
CRUXIBLE_MODE权限模式:admin(默认)、graph_write 或 read_only查看英文原文
Permission mode: admin (default), graph_write, or read_onlyCRUXIBLE_ALLOWED_ROOTS以逗号分隔的绝对路径列表,用于限制 cruxible_init 可以访问的目录。查看英文原文
Comma-separated absolute paths restricting which directories cruxible_init can access
说明为自动翻译,变量名保持原样;不确定时请查看原文。
菜单名称已于 2026年11月11日 对照 VS Code 帮助页面核实 · 你的界面可能因版本略有不同
开发者专用:终端命令和 JSON 设置
安装在电脑上
Claude Code
claude mcp add --scope user cruxible-core -- uvx cruxible-coreClaude Desktop (claude_desktop_config.json)
{
"mcpServers": {
"cruxible-core": {
"command": "uvx",
"args": [
"cruxible-core"
]
}
}
}Cursor (~/.cursor/mcp.json)
{
"mcpServers": {
"cruxible-core": {
"command": "uvx",
"args": [
"cruxible-core"
]
}
}
}VS Code (mcp.json)
{
"servers": {
"cruxible-core": {
"type": "stdio",
"command": "uvx",
"args": [
"cruxible-core"
]
}
}
}需要你填写的值
CRUXIBLE_MODEenv权限模式:admin(默认)、graph_write 或 read_only查看英文原文
Permission mode: admin (default), graph_write, or read_onlyCRUXIBLE_ALLOWED_ROOTSenv以逗号分隔的绝对路径列表,用于限制 cruxible_init 可以访问的目录。查看英文原文
Comma-separated absolute paths restricting which directories cruxible_init can access
说明为自动翻译,变量名保持原样;不确定时请查看原文。
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。
一览
哪些应用可以使用它
依据这个服务器的连接方式和各应用自己的文档,检查于 2026年11月10日。
在你的电脑上运行,适用于
ClaudeGooseLM StudioMsty StudioCherry StudioCursorVS Code + GitHub CopilotVisual StudioWindsurf (Devin Desktop)ZedKiroClineContinueJetBrains AI AssistantClaude CodeCodex CLIGemini CLIOpenCodenodecodeWarpLibreChatRaycast
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20日
- ✓ 通过
作者与仓库一致
注册名称已通过 GitHub 账户 cruxible-ai 验证,该账户拥有此仓库
- ✓ 通过
仍在维护
代码最后更新于昨天
- ✓ 通过
有许可证
使用 Apache-2.0
- – 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
已公开的漏洞(OSV.dev)
OSV.dev 中没有 cruxible-core@0.1.2 的记录
只检查这个版本的包本身,不包括所有依赖 · 检查于 2026年11月15日
作者:申请重新检查或报告信息错误
给作者:在 README 中加徽章
如果你是作者,可把这段代码贴到 README。徽章会链接回本页。
[](https://mcp.popnx.com/en/servers/cruxible-ai-cruxible-core)讨论
提问、分享技巧或报告无法使用。按“有帮助”数量排序。
用户评价
还没有评价
与“开发者工具”中的其他服务器比较i
- 星标数多于此分类中其他 6209 个服务器的 92%
- 发布到注册中心的时间比此分类中其他 6209 个服务器的 4% 更新
- 基础检查结果等于或优于此分类中其他 6209 个服务器的 99%
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
Cruxible Core 安全吗?
它在 3 项适用的基础检查中通过了 3 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 仓库最后更新于昨天,最新版本(0.1.2)于9个月前进入注册中心。
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:GitDiagram, Appllama, Cloudflare MCP, Omnara, Repowise
功能相近的其他选择
同一分类、其他作者、命令已通过检查。官方的排在前面,再按星标排序。
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
GitDiagram
Architecture diagrams, explanations and Mermaid source for public GitHub repositories.
Appllama
Study screens, flows and paywalls from top-earning iOS apps, then build from what wins.
Omnara
Deploy durable AI agents with Omnara, an open-source managed-agent platform.
Repowise
Codebase intelligence for AI coding agents: graph, git history, docs, decisions, code health.
| 本项 Cruxible Core | GitDiagram | Appllama | Cloudflare MCP | Omnara | Repowise | |
|---|---|---|---|---|---|---|
| 运行位置 | 需要安装在电脑上 | 在云端运行,无需安装 | 在云端运行,无需安装 | 在云端运行,无需安装 | 在云端运行,无需安装 | 云端或安装在电脑上都可以 |
| 需要密钥 | 否 | 否 | 否 | 否 | 否 | 否 |
| 可在 Claude 中按 + 添加 | 否 | 否 | 否 | 是 | 否 | 否 |
| 基础检查 | 3/3 | 4/4 | 4/4 | 4/4 | 4/4 | 5/5 |
| 许可证 | Apache-2.0 | MIT | MIT | Apache-2.0 | Apache-2.0 | AGPL-3.0 |
| 价格 | 登记中未列出费用 | 登记中未列出费用 | 登记中未列出费用 | 登记中未列出费用 | 登记中未列出费用 | 登记中未列出费用 |
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。
Cruxible
Cruxible is a toolset that lets AI agents create, update and manage what they’ve learned through a typed, code-like interface, so they model the world rather than solely write notes about it. Their observations and judgments become building blocks that later sessions reuse instead of re-deriving.
It runs unattended as a local daemon. Pipelines you approve re-check sources on a schedule, and when the evidence behind a fact changes, the fact is flagged instead of going quietly stale. Agents reach it through a Python SDK, an MCP server or the CLI; all three talk to the same daemon, so they share one record.
Runs on Linux (x86-64) and macOS (Apple silicon) with Python 3.11+; on Windows, use WSL2. Install and start.
A short example
A small shop promises customers delivery dates. Whether a promise is safe depends on when the supplier restocks, and the supplier's notice page changes without warning. A pipeline the shop approved re-reads that page on a schedule and keeps each reading. An agent reads the latest one and records its judgment with the Python SDK:
w = cx.world() # typed handles for this shop's own vocabulary
order = w.shop.order["ord-1042"]
the judgment: promise Friday delivery, citing the reading it rests on
order.set( promise="open", promised_for="2026-11-20", evidence=CaptureEvidence(capture="CAP-c01ffabbbb61"), # the latest reading of the notice because="Acme's notice confirms a restock of 200 for Thursday the 19th, " "so we can promise Friday the 20th.", )
-> a proposal; a second agent reviews it before it counts
A second agent reviews the proposal, and sees the supplier's words under the judgment that rests on them:
$ cruxible proposal review sha256:e0518aa5...
Why: Acme's notice confirms a restock of 200 for Thursday the 19th, so we can promise Friday the 20th. set shop.order/ord-1042 promise: open [CLM-3d57ae93...; verdict supported] cites CAP-c01ffabbbb61 acme.notice | {"restock": {"sku": "blue-widget", "quantity": 200, "date": "2026-11-19", "message": "Restock of 200 blue widgets confirmed for Thursday the 19th."}}
Then the supplier changes the page, and nobody at the shop has to notice. On its next run (every 20 seconds here) the pipeline reads the new page and records what it now says. Copying a fact from a source is the mechanical part, so it lands without review. Twelve seconds after the page changed, the promise that rested on the old date is back in front of an agent, through a saved query the shop wrote (Lines and automation):
$ cruxible get shop.item/blue-widget --detail history --field restock_expected
restock_expected by owner changed 2026-11-19 -> 2026-11-27 reason: As the supplier's restock notice states. (Procedure shop.notice-check, run RUN-8af69fc0...) by owner stated 2026-11-19 reason: As the supplier's restock notice states. (Procedure shop.notice-check, run RUN-d2498007...)
$ cruxible query --name shop.promises_at_risk subject promised_for restock_expected sku shop.order/ord-1042 2026-11-20 2026-11-27 blue-widget
The session that picks this up doesn't start over. It reads why the promise was made and what the notice says now, and changes its mind on the record, citing the new reading:
$ cruxible get shop.order/ord-1042 --detail history --field promise
promise by owner, approved by reviewer changed open -> withdrawn reason: Acme's notice now says the restock moved to Friday the 27th, so Friday the 20th no longer holds. Customer told; new date when it lands. by owner, approved by reviewer stated open reason: Acme's notice confirms a restock of 200 for Thursday the 19th, so we can promise Friday the 20th.
Why this beats notes alone
Agents today have two places to put what they learn. Notes in markdown are easy to write and hold reasoning well, but they can only be searched, not queried, and they pile up: a later note can contradict an earlier one, and nothing marks which is current. Structured data can be checked and queried, and a new value replaces the old one, but reducing everything an agent learns to fields loses detail. Research on agent memory finds both costs. In LongMemEval (ICLR 2025), storing only the facts extracted from a conversation, in place of its text, lowered accuracy, while keeping the text and indexing it with those facts beat the text alone. In MemoryAgentBench, when a later statement contradicted an earlier one, the memory systems tested often failed to answer from the newer one.
Cruxible is designed to keep the strengths of both. Facts are typed and queryable like structured data, and types can grow as agents learn more about your domain. Each fact has one current value: a new one replaces it, and the earlier values are kept as history, as the promise above shows. Reasoning stays in prose, in your own pages, and facts cite the passages they rest on. Your files, repos and services remain the source of truth: Cruxible points to them and keeps a snapshot only of the evidence a fact was drawn from, so every claim can be traced back.
- Fresh, not silently stale. Pipelines re-check facts on the schedule you set. When a source changes, the fact is updated or flagged, and a fact whose evidence is older than the window you set reads as stale instead of passing as current.
- Knowledge compounds. A fact checked once is reused by every later session and every agent until its evidence moves; work goes where the world changed.
- Governed, not only shared. Many agents and people work in one record instead of overwriting each other's notes. Every change is a proposal, checked against the rules you declared; your approval policy decides what lands at once and what waits for a second agent's signed review, and what a pipeline may write on its own is granted explicitly and expires.
- Every fact has its provenance. What was concluded, from which evidence, proposed by whom, approved by whom and when. Accepted state is a signed Git ledger, each fact keeps its own history, and any earlier value can be read back as it stood.
- Deterministic. No LLM runs inside Cruxible. Agents and people propose and judge; Cruxible applies declared rules deterministically.
Herbert Simon's watchmakers show why complex systems tend to be built from stable intermediate parts. Tempus builds each watch in one go, and every interruption makes him start over. Hora builds stable subassemblies, so an interruption costs only the piece in hand. Agent sessions with fresh context windows work like Tempus: context is pulled in on demand, and facts are re-derived and re-interpreted each time. Cruxible gives agents Hora's subassemblies. Each accepted fact has a fixed place: it belongs to the thing it describes, links to the things it relates to, and keeps the evidence it was derived from. Ask about one thing and everything known about it comes back together, so the cost of working something out is paid once and knowledge accumulates.
The security kit: the same loop on real work
A kit is a ready-made domain. The security kit keeps a project's dependencies, their known vulnerabilities and your team's judgments about them current. One Runbook reads the SBOM and asks OSV, CISA KEV and EPSS; here on a deliberately vulnerable Django app with 32 components:
$ cruxible runbook run security.start
1 Line:security.sbom-import.app written 32 (settled 32, proposed 0) 2 Line:security.osv-query written 181 (settled 181, proposed 0) 3 Line:security.kev-triage written 181 (settled 181, proposed 0) 4 Line:security.epss-triage written 543 (settled 543, proposed 0) 5 Line:security.exposure-attention.unjudged inbox 7 new
Facts copied from sources settle without review. The seven inbox items are unjudged exposures above the kit's exploit-likelihood threshold, and judging one is the agent's work: it reads the code, records its verdict with the passage it rests on, and a second agent reviews it.
$ cruxible get security.app_exposure/sbom.cdx.json--pkg-pypi-pyyaml-5.1--cve-2020-14343 --detail history
justification proposed 2026-11-18 by owner, approved by reviewer stated vulnerable_code_not_in_execute_path reason: CVE-2020-14343: the only yaml.load in the repo is in dockerized_labs/a9_uckv_lab, a separate container with its own requirements; the Django app this SBOM describes never imports yaml.
When PyYAML is upgraded and the Runbook runs again, only what changed is written; everything else is re-checked and left as it is:
$ cruxible runbook run security.start
1 Line:security.sbom-import.app written 1; unchanged 31; retired 1 2 Line:security.osv-query written 1; unchanged 178 3 Line:security.kev-triage written 0; unchanged 179 4 Line:security.epss-triage written 0; unchanged 537
The security walkthrough goes through all of it on your own project. cruxible console opens a read-only view of the same state in your browser.
Install and start
Requirements: Python 3.11+ and Git. Linux x86-64 and macOS on Apple silicon; on Windows, use WSL2.
pip install cruxible # or: uv tool install cruxible
cd your-project cruxible init --principal-id owner \ --reviewer-key-dir ../reviewer-keys --require-independent-approval cruxible orient # what this instance holds cruxible next # what needs attention
init starts a local daemon if none is running and creates this project's instance. The two reviewer options set up a second principal whose approval every proposal needs; cruxible init --principal-id owner alone lets one owner approve everything. The Quickstart continues from here with the security kit.
Use it from your agent
One daemon owns the state; three clients talk to it:
- MCP: an MCP client launches
cruxible mcp(uvx cruxible mcpfrom the registry listing). See MCP tools. - CLI: every command above, for agents with a shell and for people.
- Python SDK:
cruxible-client, installed with the package. It reads your vocabulary into typed Python, as in the example at the top. See the SDK reference.
The verbs are the same on all three: orient, query and get read; set and write change values; authoring changes definitions. init starts a local daemon for you; cruxible server start runs one yourself.
The plugin for Claude Code and Codex adds five general skills: survey works out what in a project with no kit is worth holding as state, adopt-kit starts from a kit, and create-state, automate-state and review-state cover the work after that. For AI agents covers how an agent should read and write.
How it works
- Facts, vocabulary and evidence. Each fact is a Claim about a Subject, under a ClaimType that says what the value may be and what may back it: the author's own words, a cited passage of a file, or a Capture (the slice of a