semaphore-mcp
A curated MCP server for Semaphore UI that enables driving Ansible and Terraform automation from Claude with a safe, read-first tool surface.
README
semaphore-mcp
A curated MCP server for Semaphore UI — drive your Ansible / Terraform automation from Claude with a safe, tiered tool surface: read freely, write deliberately, delete only with an explicit confirmation.
The token never passes through Claude's context — it's an environment variable on your machine. The exposed tool set is generated from Semaphore's own OpenAPI spec and guardrails deletes rather than pretending they don't exist.
Why generated, and why curated
Semaphore's API is 127 operations and ships empty operationIds, so a generic OpenAPI→MCP bridge produces unusable tool names.
generate.py solves that: it reads the pinned api-docs.yml, filters to a hand-picked allowlist, synthesises clean tool names, and derives readOnlyHint / destructiveHint annotations from each tool's declared tier. The result is 27 tools across three tiers:
| Tier | Count | readOnlyHint |
destructiveHint |
Guard |
|---|---|---|---|---|
| read | 12 | True |
False |
safe to auto-allow |
| write | 10 | False |
False |
ask — state-changing, non-destructive |
| delete | 5 | False |
True |
ask + refuses unless confirm=True |
Covered resources: projects, templates, tasks, inventories, environments (vars), and repositories (create/update/delete — needed for branch-pinned testing).
GET /project/{id}/templates -> list_templates (read)
GET /project/{id}/repositories -> list_repositories (read)
...
POST /project/{id}/tasks -> run_task (write)
POST /project/{id}/templates -> create_template (write)
PUT /project/{id}/environment/{eid} -> update_environment (write — set vars)
POST /project/{id}/repositories -> create_repository (write — pin a branch)
...
DELETE /project/{id}/templates/{tid} -> delete_template (delete, guarded)
DELETE /project/{id}/tasks/{tid} -> delete_task (delete, guarded)
Two safety properties worth calling out:
- Deletes refuse on the first call. Every
delete_*returns a refusal payload unless invoked withconfirm=True, so a stray delete can never fire by accident. - Updates never blank what you didn't touch.
update_*tools are read-modify-write: theyGETthe current object, overlay only the fields you pass, thenPUTthe merged result — so renaming a template can't silently wipe its arguments or survey vars. (Environment secrets are write-only in the API and are never echoed back; omitting them leaves stored secrets intact.)
Widen or narrow the surface by editing the ALLOWLIST in generate.py and regenerating.
Quick start (Docker)
Pull the pre-built image from GitHub Container Registry:
docker pull ghcr.io/gabrielbelli/semaphore-mcp:latest
Mint a scoped API token in Semaphore (User settings → API Tokens) on an RBAC user scoped to match the tier you intend to allow — the token is the last, hardest fence. Then register with Claude Code by adding this to ~/.claude/mcp.json:
{
"mcpServers": {
"semaphore": {
"command": "docker",
"args": [
"run", "--rm", "-i",
"-e", "SEMAPHORE_URL=https://semaphore.example.com",
"-e", "SEMAPHORE_TOKEN=your-scoped-token",
"ghcr.io/gabrielbelli/semaphore-mcp:latest"
]
}
}
}
-i (not -t) is required for stdio MCP transport. Restart Claude Code, then call list_projects() to verify.
First publish: GHCR images inherit the source repository's visibility only after the image is linked. The
org.opencontainers.image.sourcelabel is set automatically by the workflow; on first publish, visit Your profile → Packages → semaphore-mcp → Package settings and either link it to the repo or flip visibility to public.
Four gates against a destructive call
| Layer | Enforced by | Real fence? |
|---|---|---|
destructiveHint / readOnlyHint annotations |
MCP client (advisory) | hint only |
confirm=True gate on every delete_* |
the server itself (in-process) | ✅ |
Claude Code ask / deny rules |
Claude Code permissions | ✅ |
| Scoped Semaphore API token | Semaphore RBAC (returns 403) |
✅✅ |
Belt-and-braces permission block for Claude Code settings.json — reads run silent, writes prompt, deletes are denied outright at the client (drop them into ask instead if you want to allow them):
{
"permissions": {
"allow": [
"mcp__semaphore__list_projects", "mcp__semaphore__get_project",
"mcp__semaphore__list_templates", "mcp__semaphore__get_template",
"mcp__semaphore__list_tasks", "mcp__semaphore__get_last_tasks",
"mcp__semaphore__get_task", "mcp__semaphore__get_task_output",
"mcp__semaphore__list_inventory", "mcp__semaphore__get_inventory",
"mcp__semaphore__list_environment", "mcp__semaphore__list_repositories"
],
"ask": [
"mcp__semaphore__run_task", "mcp__semaphore__stop_task",
"mcp__semaphore__create_template", "mcp__semaphore__update_template",
"mcp__semaphore__create_environment", "mcp__semaphore__update_environment",
"mcp__semaphore__create_inventory", "mcp__semaphore__update_inventory",
"mcp__semaphore__create_repository", "mcp__semaphore__update_repository"
],
"deny": [
"mcp__semaphore__delete_template", "mcp__semaphore__delete_environment",
"mcp__semaphore__delete_inventory", "mcp__semaphore__delete_repository",
"mcp__semaphore__delete_task"
]
}
}
Local development
python3 -m venv .venv && ./.venv/bin/pip install -r requirements.txt
./.venv/bin/python generate.py # api-docs.yml -> server.py
./.venv/bin/python server.py --list # list tools + annotations (offline)
./.venv/bin/pytest -q # safety-surface tests (offline)
Or via make: make list, make build, make test, make run.
Bumping the pinned spec on a Semaphore upgrade
curl -sL https://raw.githubusercontent.com/semaphoreui/semaphore/develop/api-docs.yml -o api-docs.yml
python3 generate.py && git diff --stat
Licence
BSD 2-Clause. 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。