Taproom
Discovers and manages portable agent capabilities (skills and MCP servers) from configurable collections, providing search, inspection, and local installation via CLI and MCP tools.
README
Taproom

Taproom discovers portable agent capabilities without owning them. It indexes skill collections and MCP server registries, exposes skills through standard FastMCP resources, and provides a local CLI for explicit installation.
The development defaults discover sibling checkouts of the public SkillTap and
MCP Tap collections. A taproom.toml file can register any number of named
public or private taps, each containing multiple skill and MCP sources. Private
collections are never loaded implicitly.
Architecture
- The hosted FastMCP service searches and serves capability metadata.
- FastMCP's skill provider exposes a namespaced
skill://resource, generated manifest, and supporting files for every indexed skill. - A small MJS CLI performs local filesystem changes, verifies SHA-256 hashes,
and records installed files in
.taproom.lock. - MCP packages are searchable, inspectable, hash-verifiable, and fetchable. Automatic activation requires declared host requirements and a passing local doctor check.
Quickstart
Requires Python 3.11+, uv, and FastMCP 3.4.x.
git clone https://github.com/batteryshark/taproom.git
git clone https://github.com/batteryshark/skills-tap.git skill-tap
cd taproom
uv sync
uv run taproom-server
The MCP endpoint is http://127.0.0.1:8768/mcp. The CLI-facing HTTP API is
under http://127.0.0.1:8768/api/v1/.
Configure taps
Copy taproom.example.toml to the ignored local file
taproom.toml, then enable and edit the taps available on that host:
cp taproom.example.toml taproom.toml
Each tap has a stable name and visibility. Sources inside it have their own names, kinds, and roots:
version = 1
[[taps]]
name = "private"
visibility = "private"
[[taps.sources]]
name = "skilltap"
kind = "skill"
path = "../skill-tap-private/skills"
[[taps.sources]]
name = "rekit"
kind = "skill"
path = "../../reverse-engineering-tools/rekit-private/skills"
[[taps.sources]]
name = "rekit"
kind = "mcp"
path = "../../reverse-engineering-tools/rekit-private/mcp-servers"
Relative source paths resolve from the config file. Set TAPROOM_CONFIG to use
a config stored elsewhere. The older TAPROOM_SKILL_ROOTS and
TAPROOM_MCP_ROOTS variables remain available for a single compatibility tap.
Set TAPROOM_HOST or TAPROOM_PORT to override the default bind address.
The visibility field is descriptive metadata; it does not enforce access.
Remote deployment
Taproom binds to 127.0.0.1 by default. Its CLI-facing HTTP routes do not
provide built-in authentication. Put a remotely bound instance behind an
authenticated reverse proxy, and set TAPROOM_SKILL_ROOTS and
TAPROOM_MCP_ROOTS or TAPROOM_CONFIG to an explicit allowlist. Never expose
private collection roots through an unauthenticated deployment.
Use the local CLI
The CLI has no npm dependencies and requires Node.js 20+.
node bin/taproom.mjs taps
node bin/taproom.mjs search "understand an inherited codebase"
node bin/taproom.mjs search "trace a native process" --tap private
node bin/taproom.mjs info public:skill:skilltap~development~codebase-archeology
node bin/taproom.mjs add public:skill:skilltap~development~codebase-archeology --project .
node bin/taproom.mjs add public:skill:skilltap~development~codebase-archeology --global
Capability IDs use tap:kind:source~path, so identical skill names from
different taps, sources, or categories remain independently addressable.
MCP packages
Inspect requirements before fetching or activating a server:
node bin/taproom.mjs plan public:mcp:mcptap~travel~sandals
node bin/taproom.mjs doctor public:mcp:mcptap~travel~sandals
node bin/taproom.mjs fetch public:mcp:mcptap~travel~sandals --project .
node bin/taproom.mjs configure public:mcp:mcptap~travel~sandals --project . --client codex
plan reports runtime, transport, launch configuration, required environment,
declared host requirements, detected Python or Node dependencies, lockfiles,
and unresolved setup. fetch downloads the complete package and verifies every
file hash without installing system dependencies. doctor checks the current
platform, commands, and required environment variables. configure prints a
reviewable client configuration; it does not modify client settings.
add combines doctor and fetch. It refuses MCP activation when requirements
are incomplete or the local checks fail. An explicit fetch remains available
after the operator reviews those warnings.
For dependency-aware activation, an MCP package's server.json declares host
requirements alongside its existing runtime, launch, and environment fields:
{
"requirements": {
"platforms": ["windows"],
"commands": [
{"name": "uv", "version": ">=0.10"}
],
"software": [
{"name": "x64dbg", "version": "2026.05.27"}
],
"setup": [
"Copy the matching plugin into the x64dbg or x32dbg plugin directory."
]
}
}
Supported platform values are any, linux, macos, and windows.
commands are checked on PATH; software and setup are presented as
manual verification items. If requirements is absent, Taproom reports the
manifest as incomplete and points the operator to the included README and
dependency files.
Project installs go to <project>/.agents/skills/. Global installs default to
~/.agents/skills/; set TAPROOM_SKILL_HOME to target another agent's skill
directory. Existing skills are never replaced unless --force is supplied.
MCP tools
list_taps()search_capabilities(query, kind?, tap?, limit?)inspect_capability(capability_id)plan_mcp_server(capability_id)
Skills are also available through FastMCP's standard skill resources. Search is implemented over capability metadata because FastMCP's BM25 search transform indexes tools rather than resources.
Repository map
src/taproom/contains catalog discovery, ranking, and the FastMCP service.src/taproom/packages.pybuilds file manifests and MCP dependency plans.bin/taproom.mjsis the dependency-free local installer.tests/covers catalog identity, manifests, search, and CLI installation.assets/contains the project artwork.
Test
uv run python -m unittest discover -s tests
npm test
License
MIT. See 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。