personal-repo-mcp
Enables AI agents to work with administrator-approved Git repositories on a self-hosted VPS, supporting cloning, file edits, and full Git operations within persistent workspaces.
README
personal-repo-mcp
A self-hosted, multi-repository MCP server for AI agents. It gives an agent a persistent Git workspace on your own VPS, with controlled access to administrator-approved repositories and normal Git operations.
Version 1.0.0 — MVP
Core idea
The administrator controls what repositories the agent may access. The agent controls which allowed repositories it actually clones.
Administrator
|
| mcp-config add repo jsilvanus/*
v
repositories.json (allow-list)
|
v
MCP server
|
+-- clone_repository("jsilvanus/project")
|
v
persistent workspace
The workspace is an ordinary Git repository. GitHub remains an upstream remote; the MCP server is the persistent local workspace and control boundary.
MCP help resources
The server provides modular, MCP-native operational documentation. Start with:
mcp://help/index
Focused topics are available at:
mcp://help/repositories
mcp://help/files
mcp://help/git
mcp://help/chain-command
mcp://help/resources
The help resources are intentionally separate from tools/list: the MCP tool list remains the authoritative API surface, while the help resources provide progressive-disclosure workflow guidance for agents.
Features
- Persistent multi-repository workspaces
- Administrator-controlled repository allow-list
- GitHub PAT authentication for Git operations
- Streamable HTTP transport
- File reads, line-range reads, writes, line edits, search, and patches
- Git status, diff, history, branches, fetch/pull/push, merge/rebase, and conflict handling
- Single-repository
chain_commandfor composing operations - MCP resources for system, repository, Git, and individual file state
- Resource subscriptions and change notifications
- Secret scrubbing from MCP output
- Git submodule safeguards
- Docker deployment
- Optional hot-git backend for persistent, treeless repository reads and edits
Git backends
The existing Git command backend remains the default compatibility path.
The new hot-git backend can be enabled with:
export PERSONAL_REPO_MCP_GIT_BACKEND=hot-git
The dependency is declared in pyproject.toml and is installed as a normal Python package; it is not a Git submodule.
The backend boundary is deliberately small:
MCP tool
|
v
Repository backend
|
+-- Git backend
|
+-- hot-git backend
|
+-- persistent object reader
+-- treeless object/tree/commit edits
+-- Git CAS ref publication
When hot-git is selected, each repository gets one lazily-created long-lived backend instance. It is shared by callers and closed with the server. This keeps the persistent cat-file --batch reader alive without creating one process per MCP request.
The first integrated operations are:
git_read_file— read a file from a ref or commit through the configured backend;git_edit— perform a multi-file treeless edit and publish it with an expected-ref CAS.
The existing Git command tools remain available during the migration. Operations that depend on a working tree or Git's higher-level merge/rebase machinery continue to use the compatibility backend.
Chained edits
git_edit returns the resulting commit. A caller can use that commit as the expected base for the next edit:
A(base=X) -> commit A
B(base=A) -> commit B
C(base=B) -> commit C
This is the foundation for integrating the backend with chain_command without allowing concurrent writers to silently overwrite each other.
Deployment
Configure the host with mcp-config, then start the single Docker deployment:
mcp-config pat
mcp-config add repo jsilvanus/*
docker compose up -d --build
mcp-config add repo changes only the administrator allow-list. It does not clone repositories. The agent clones an allowed repository through the MCP when needed.
For hot-git mode, the normal Python dependency installation installs the library. No submodule checkout is required.
See the deployment documentation in docs/ for persistent storage, credentials, HTTPS, and configuration details.
License
EUPL-1.2
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。