MCP Skill Hub 87f5cfdd

MCP搜索检索

提供方 ccancellieri

搜索检索 stdio · 本地进程 Token 社区 / 第三方 Claude
查看接入方式

MCP Skill Hub是一个本地MCP服务器,通过语义技能搜索、零令牌钩子拦截和四级本地执行引擎,使Claude Code更智能、更经济且具备离线能力。

一句话选型结论

按资料字段整理

MCP Skill Hub 87f5cfdd 属于搜索检索类 MCP 服务器,stdio · 本地进程,Token,可直接接入 Claude,有可直接执行的安装命令。接入前先确认权限范围;涉及写入、删除或付款的操作,建议先在测试环境验证。

分类搜索检索 传输方式stdio · 本地进程 鉴权Token 部署方式本地部署
GitHub 星标 0 仓库热度,无公开仓库时显示未标注
工具数 未标注 暴露给模型的工具数量
资源 / 提示词 0 / 0 资源与提示词模板数量
维护方 社区 / 第三方 核验等级 仅资料核对

工具与能力

搜索检索
数量来自来源目录的标注,具体工具清单以官方文档为准
工具数量未标注暴露给模型可调用的工具个数
资源数量0可读取的资源条目
提示词数量0内置提示词模板数量

能力标签

支持客户端

适用场景

语义搜索Python本地部署本地执行离线能力零令牌拦截技能管理

安装与配置

stdio · 本地进程 Token
客户端在本机拉起子进程,数据不出本机
python install.py # cross-platform

这条命令只是入口,具体启动参数请按官方文档补齐,不要直接照搬到别的服务器。

  1. 1
    准备运行环境

    stdio 服务在本地起进程,需要先装好 Python,再按安装命令拉取依赖。

  2. 2
    安装服务

    复制上面的安装命令执行;命令里带占位符的位置,换成自己的密钥或路径。

  3. 3
    写入客户端配置

    在客户端的 MCP 配置里填好启动命令和参数,保存后完全退出客户端再打开,让新配置生效。

  4. 4
    验证是否可用

    连接成功后随便调一个只读工具,确认能返回结果,再接入正式流程。

典型使用场景

3 类
按分类与标签整理的常见用法,具体以官方文档为准
联网搜索取证

让模型查最新资料并给出出处。

资料聚合

把一个主题的网页、文档集中抓回来整理。

情报与竞品检索

按关键词持续检索行业动态和竞品信息。

选择建议

选之前先看这几条
按条目字段给出的判断,不替你决定用不用

值得选的地方

  • 有现成的安装命令,不用自己找入口,照着跑就能起服务。
  • 已适配 Claude,换客户端不用重新研究配置。

先想清楚的地方

  • stdio 方式要求本地装好运行时(如 Node.js),换机器要重新装一遍。
  • 使用前要准备好Token对应的凭据,并按最小权限配置。
  • 资料没标注工具数量,实际能力要装好后确认。
  • 版本更新与行为细节以官方仓库为准。

更适合

  • 需要在搜索检索场景里让模型直接调用外部能力的人。
  • 已经在用 Claude 的用户,接入路径最短。
  • 想先快速试一下、能接受在本地或自有环境跑服务的人。

不太适合

  • 完全不想在本地装环境的人。
  • 不方便管理密钥或第三方授权的人。
  • 需要厂商级 SLA 和技术支持的生产环境。

同类对比

6 个同类
同分类里挑最接近的几条,字段以各自详情页为准
MCP 服务器工具数传输方式部署方式安装命令最后核验
MCP Skill Hub 87f5cfdd当前 未标注 stdio · 本地进程 本地部署 有 2026-05-17
Brave Search 433812ff 2 未标注 未标注 无 2026-05-17
Sequentialthinking 1 SSE · 服务端推送 远程部署 有 2026-05-18
Baidu Netdisk 未标注 未标注 未标注 无 2026-05-17
ragflow 未标注 stdio · 本地进程 本地部署 有 2026-04-11
mindsdb (430e18) 未标注 未标注 未标注 有 2026-05-17

服务器档案

搜索检索
字段来自项目仓库与公开索引,中文说明由本站标注
提供方ccancellieri上游标注的维护方 ccancellieri
分类搜索检索本站按上游分类展示
连接方式stdio · 本地进程客户端在本机拉起子进程,数据不出本机
鉴权方式Token远程服务建议用最小权限凭据
维护方社区 / 第三方核验等级 仅资料核对
包名 / 运行时未标注 · Python本地安装时用到的包标识与运行时
部署类型本地部署上游给出的部署形态
最近核验2026-05-17上游最后核验时间

详细介绍

7.5 KB
来自项目仓库与上游目录的原文,未做删改

MCP Skill Hub

Alpha · Apache License 2.0 · Author: Carlo Cancellieri

Choose the evidence a coding assistant needs, then keep the context small.

MCP Skill Hub is a local Python toolkit for project memory, skill discovery
and reviewable context composition. It started as an MCP server to provide a
standard integration interface across AI clients and agent runtimes
("harnesses"). The name reflects those origins; today the project exposes
several interfaces:

  • MCP tools for clients that support the Model Context Protocol.
  • CLI tools and a shared JSON interface for scripts and direct client integrations.
  • Plugin extension points for adding web views, storage and indexing capabilities.
  • Optional hooks and native adapters for supported client events.
  • A web dashboard and context composer for inspecting sources, reviewing context and copying it into a client.

MCP is one way to access these capabilities. Hooks and native adapters are
client-specific; choose the integration supported by your harness.

The core retrieval path uses SQLite and keyword search: no model download or API
key is needed. The original prompt stays intact.

Get started · Context composer ·
Measured results · Client integrations

What works today

Capability What you can do Verification
Minimal MCP profile Expose eight context tools; load full skill instructions only when needed. Protocol tests verify the advertised tools, reject excluded calls and prevent background-service startup.
Manual context composer Select projects, review candidate sources, choose exact passages, preview and copy a bounded packet. UI and service checks cover selection, expansion, preview and source changes.
Scoped memory retrieval Retrieve project evidence only from explicitly supplied project roots. Missing scope allows global skills only. Synthetic fixtures and scope regressions check foreign-source rejection.
Deterministic compression Compact JSON whitespace while preserving strings, duplicate keys and numeric spelling; identify lossy excerpt selection and truncation. Fidelity tests cover duplicate members, huge exponents, escapes, Unicode and code.
Source recovery Expand a selected candidate or fetch a skill's full indexed text. Changed sources require a new selection. Fingerprint and stale-source tests; retained originals are used instead of generated digests.
Optional client adapters Add bounded evidence while preserving the prompt and native client approvals. MCP and manual copy work without hooks. Python and Node adapter tests; native-client qualification is tracked separately.

The full MCP profile remains the compatibility default. The minimal profile is
an explicit choice and does not configure providers, install hooks or start a
model service. Claude, Codex, Pi and OpenClaw integration paths are documented;
adapter tests do not imply every live client configuration has been qualified.

Measured, with limits

Comparison Result What was measured
Full → minimal MCP surface 87 → 8 tools; 92.41% fewer serialized tokens Tool schemas and server instructions: 16,362 → 1,242 o200k_base tokens, counted once as compact JSON.
Project isolation 0 labeled foreign-source leaks across 144 cases Frozen synthetic project-scope fixtures, not an audit of every historical memory record.

These measurements concern protocol and context payloads. A reduction in total
tokens per completed task has not been demonstrated.
Clients may cache or defer
tool definitions, and extra context can increase usage. See the
reproducible report for versions, commands,
corpus scope and limitations, and the task evaluation protocol.

Local selector studies did not justify promotion. Model-based ranking,
preference learning and automatic prompt rewriting are not part of the normal
composer workflow. A more complex model must earn its place through task-level
results.

Get started

Requires Python 3.11+. For the core workflow, install the Python package:```sh
git clone https://github.com/ccancellieri/mcp-skill-hub.git
cd mcp-skill-hub
python3 -m venv .venv
.venv/bin/python -m pip install -e .


On Windows, use the executables in `.venv/Scripts` instead of `.venv/bin`.
The package installation does not register hooks or download model weights.
Index a skill directory you explicitly choose, without generating embeddings:```sh
.venv/bin/skill-hub-cli index_skills_text --skill-dir /absolute/path/to/skills

Repeat --skill-dir for additional directories. The command scans their
SKILL.md files only. It does not scan other configured locations or index
project memory. See MCP profiles for the tool list.

Configure your MCP client's stdio server with the installed executable and
explicit minimal profile. A typical server entry is:```json
{
"command": "/absolute/path/to/mcp-skill-hub/.venv/bin/skill-hub",
"args": ["--profile", "minimal"]
}


Use your client's configuration format and reconnect after changing it.
The server uses the existing local index; select and index your skill sources
before searching. Installation options cover the full
installer and optional model-backed services.

## Use only the context you need

1. **Discover:** call `search_skills` for short descriptions; use
   `get_skill_content` when a specific skill is needed.
2. **Retrieve:** call `prepare_context` with the original request and the
   absolute project path. Retrieved text is evidence, not an instruction or
   authorization.
3. **Review:** use `prepare_composition`, `expand_context_candidate` and
   `compose_context`, or open the dashboard's `/context` page. Review the sources,
   budget and any lossy transformations before copying the result.```python
# MCP tool arguments, sent through your client:
prepare_context(text="Review the migration constraints", repo_root="/path/to/project")
search_skills(query="database migration")

For the web composer, run skill-hub-dashboard --no-services separately and open
http://127.0.0.1:8765/context. This option skips background service startup and
maintenance; it does not stop services already running. Starting the minimal MCP server does
not launch the dashboard. Token counts in the
composer are text estimates. Optional prompt compression is an explicit
original/proposal/diff action and never replaces the request automatically.

Boundaries and compatibility

  • Project scope is explicit. Neither ranking history nor client identity grants
    access to another project's memory.
  • The automatic context hook is deterministic, has a two-second adapter deadline
    and makes no L1/L2 model call. Failure adds no context; the prompt passes through.
  • Client model and effort observations are separate from the server's configured
    providers. Unreported runtime metadata remains unknown.
  • Generic auto-proceed, task interception, inferred approvals and hook-driven
    model switching have been retired. See the migration guide.
  • Optional semantic search, provider services and research APIs are documented
    separately. They are not required to use the core context service.

Development and evidence

Run the isolated Python suite and client-adapter checks:sh uv run --with pytest --with pytest-asyncio --with pytest-timeout \ python -m pytest -m 'not local_only' node --test integrations/tests/context-adapters.test.mjs

Development guide ·
Documentation index · Roadmap ·
Open issues

License

Copyright © 2026 Carlo Cancellieri. Licensed under the
Apache License 2.0.

常见问题

8 问
答案由本站依据条目字段整理,不引用上游原文
MCP Skill Hub 87f5cfdd 怎么安装?

在客户端按下面的安装命令接入:python install.py # cross-platform。装完在客户端配置里引用同一个包或端点。

MCP Skill Hub 87f5cfdd 怎么在 Claude 里配置?

它的传输方式是stdio · 本地进程,在客户端的 MCP 配置里填好启动命令和参数,保存后完全退出客户端再打开,配置才会生效。

MCP Skill Hub 87f5cfdd 收费吗?

本站不代收费用,也没有标注具体价格。是否收费取决于服务本身和背后的接口额度,正式使用前按官方说明确认。

MCP Skill Hub 87f5cfdd 需要 API Key 吗?

资料标注需要Token认证,使用前准备好对应凭据,并按最小权限配置。

MCP Skill Hub 87f5cfdd 支持哪些客户端?

资料标注支持 Claude;其他客户端只要能配置 MCP Server,也能用同样的命令或地址接入。

MCP Skill Hub 87f5cfdd 能提供哪些能力?

资料显示能力标签包括 搜索、Python、Claude、技能管理。具体工具清单以官方文档为准。

MCP Skill Hub 87f5cfdd 需要什么运行环境?

运行时为 Python;属于本地进程,需要在本机装好运行环境。

MCP Skill Hub 87f5cfdd 是本地还是远程?

它是本地进程(stdio),由客户端在本机拉起。传输方式为stdio · 本地进程。

热门专题

按标签挑服务器
全部专题 →

服务器信息以作者仓库与官方文档为准;分类、传输方式与鉴权方式的中文说明为中文整理,最近核验 2026-05-17。