@humanagencyp/deploy-mcp
Enables agents to deploy software via GitHub Actions only with a signed receipt, providing gated tools to resolve refs, list deployments/environments, and dispatch pipelines for a specific commit.
README
@humanagencyp/deploy-mcp
Deploy MCP server — run deployment pipelines under bounded human authority.
Implements the executor side of the Human Agency Protocol
deploy@0.6 profile. Backend: GitHub Actions.
What it does
Exposes deployment as gated tools so an agent can ship software only with a signed receipt behind it. It does not decide whether a deploy is allowed — that is the Gatekeeper's job, and then the pipeline's.
| Tool | Kind | |
|---|---|---|
resolve_ref |
read | branch/tag → commit SHA |
list_deployments |
read | recent workflow runs |
get_deployment |
read | one run |
list_environments |
read | the repo's real environments |
deploy |
consequential | dispatch a pipeline for one commit |
Setup
GITHUB_TOKEN=<fine-grained token>
Fine-grained, scoped to one repository:
| Permission | Why |
|---|---|
| Actions: read & write | dispatch the workflow |
| Environments: read | list environments for the authorization wizard |
| Contents: read | resolve a branch to a commit |
Not Contents: write. This server ships commits that already exist; it never
authors them.
Two properties that are easy to lose
1. deploy declares receipt_id. That declaration is what makes a Suveren
gateway inject the receipt it just minted. Remove it and the pipeline has
nothing to verify — the chain quietly degrades to an unproven deploy. The tool
refuses to dispatch when the field is missing, so the failure is loud.
2. A commit is required, not a branch. deploy rejects main. A receipt
binds one commit; authorising "deploy main" would authorise whatever main
happens to be when the action finally runs, which may not be what was reviewed.
Call resolve_ref first.
The pipeline must verify
Dispatching is not the control point — anyone with repository write access can dispatch a workflow. The workflow itself must check the receipt before releasing anything:
on:
workflow_dispatch:
inputs:
sha: { required: true }
receipt_id: { required: true }
environment: { required: true }
jobs:
verify: # signature, scope, commit binding, freshness
...
deploy:
needs: verify
steps:
- uses: actions/checkout@v6
with:
ref: ${{ inputs.sha }} # the APPROVED commit, not the branch head
Building the branch head instead of inputs.sha means a receipt for one commit
can ship another — and the receipt still verifies, certifying something false.
That is worse than having no receipt at all.
Limits, stated plainly
- This server dispatches; it does not hold the deploy credentials the pipeline uses. The boundary is the receipt requirement, not the whole path.
- GitHub returns no run id from a dispatch, so the run is matched by workflow and creation time. Best-effort: it reports no run rather than guessing wrong.
Licence
MIT
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。