beads-mcp-server
An MCP server that exposes a curated typed surface over the Beads CLI, managing workspaces and issues through secure, sessionless HTTP with static bearer authentication.
README
Beads MCP Server
An owned Python MCP server built on the official MCP SDK v2. It exposes a
curated typed surface over bd and keeps bd as the Beads domain authority.
Architecture
MCP SDK v2 Streamable HTTP
-> static bearer verification
-> typed tools with mandatory workspace_id
-> explicit workspace registry
-> one bounded async bd subprocess per call
-> workspace-configured shared Dolt server
The server does not import or wrap the upstream beads-mcp package. It never
queries Dolt directly and does not reproduce Beads business logic. Each call
starts a fresh bd process using an argv list, a hard timeout, bounded stdout
and stderr, cancellation cleanup, and an explicit child-environment allowlist.
Protocol sessions are not application state. Modern 2026-07-28 requests are
sessionless and every workspace-sensitive call carries workspace_id. The
registry resolves that stable identifier to an operator-controlled absolute
path; clients cannot submit filesystem paths.
The first release deliberately has no request queue, semaphore, per-workspace
lock, connection pool, or Dolt adapter. Sparse traffic from 40 to 50 connected
agents should be measured before adding coordination that may duplicate what
bd and the shared Dolt server already provide.
Tool surface
The server exposes workspaces, workspace_status, ready, list, show,
create, claim, update, close, reopen, dep, comment, comments,
note, stats, and blocked. workspaces returns only exact IDs and never
server paths. workspace_status performs a read-only bd stats --no-activity
probe for one ID. There is no mutable context tool, cwd discovery, arbitrary
command passthrough, workspace provisioning, or direct Dolt operation.
Configuration
Copy .env.example to the service's existing protected environment file and
set values through the operator's secret-management process. The service fails
closed when the bearer token, public URL, allowed hosts, or workspace registry
is absent.
For a central service with many repositories, point the server at an operator-controlled root:
BEADS_WORKSPACE_ROOT=/srv/beads-workspaces
BEADS_READINESS_WORKSPACE_ID=project-a
Every immediate child containing .beads/metadata.json or
.beads/config.yaml becomes one exact workspace_id. The server never scans
components of a client path and never treats dashes and underscores as aliases.
Consequently, library, cognovis_core, and cognovis-core can coexist
without ambiguity. An explicit BEADS_WORKSPACES_JSON object remains available
for installations that prefer a fixed mapping; configure exactly one source.
Only these variables can reach bd when present: PATH, HOME, USER,
LOGNAME, LANG, LC_ALL, BEADS_ACTOR, BEADS_DOLT_SERVER_HOST,
BEADS_DOLT_SERVER_PORT, BEADS_DOLT_SERVER_USER, BEADS_DOLT_PASSWORD, and
DOLT_PASSWORD. The server never logs their values.
The public /health route is a process liveness check. /ready performs one
configured read-only workspace probe and returns no workspace ID or filesystem
path. The /mcp route requires the configured static bearer token. Host and
Origin validation is enforced by the SDK's TransportSecuritySettings.
Local development
Python 3.14 and uv are required.
uv sync --frozen
uv run pytest -q
uv run ruff check .
uv run ruff format --check .
uv build
Run the service with an operator-provided environment:
uv run beads-mcp-server
Deployment boundary
This repository contains the reusable package rather than a specific server's
deployment. Install with uv sync --frozen --no-dev and run:
.venv/bin/beads-mcp-server
Keep service-manager units, reverse-proxy rules, concrete hostnames, secret delivery, and workspace locations in the consuming infrastructure repository. The server does not copy an environment file, create a token, initialize a Beads database, or rewrite workspace metadata.
Compatibility evidence
The test suite uses the official in-memory SDK client for modern discovery and
typed tool calls. HTTP integration tests exercise the real ASGI transport,
bearer rejection, Host and Origin rejection, and independent 2026-07-28
requests without Mcp-Session-Id.
No handshake-era client behavior is promised. stateless_http=True therefore
disables legacy back-channels; this server has no server-initiated requests or
legacy notifications that require them. Official final-spec conformance is run
against a local HTTP endpoint before release and is reported separately from
unit and client compatibility evidence.
License
Released under the MIT 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。