RepoPrimer
Provides fresh project context to coding agents by combining Markdown documentation and live Git state, enabling deterministic startup briefs and bounded document retrieval for MCP-compatible tools.
README
RepoPrimer
English · 简体中文
Fresh project context for every coding agent.
RepoPrimer is a local-first, read-only project handoff layer. It combines the Markdown documentation you already maintain with live Git state, then exposes a compact startup brief to Codex, Claude, Cursor, and other MCP-compatible coding agents.
<p align="center"> <img src="https://raw.githubusercontent.com/repoprimer/repoprimer/main/.github/assets/demo.svg" alt="repoprimer context output: a compact brief with project state, tasks, decisions, and live Git status" width="720"> </p>
One repoprimer context call returns a bounded startup brief — project state,
open tasks, accepted decisions, and live Git status. On this repository the
compact brief measures about 3 KB (≈800 tokens, within the 6,000-character
default budget) instead of a multi-turn cold-start exploration.
Status: early development. The npm package has not been published yet. The package commands below describe the intended public interface; use the source workflow while developing locally.
Why RepoPrimer?
Coding agents repeatedly spend time rediscovering the same facts: what the project does, what changed recently, what is in progress, and which decisions must not be revisited. General-purpose AI memory products solve a broader problem. RepoPrimer deliberately solves one narrow one:
Before an agent starts work, tell it where this repository is now.
RepoPrimer is designed to be:
- Local-first: project files and the registry stay on your machine.
- Read-only at runtime: MCP tools do not modify project files or Git state.
- Markdown-native: no database, embeddings, or proprietary storage format.
- Live: the handoff includes current Git state rather than only cached notes.
- Small: four MCP tools, bounded responses, and no model API dependency.
- Portable: one context source can serve multiple MCP-compatible agents.
How it works
existing Markdown docs + live Git state
|
v
compact project context
|
v
Codex / Claude / Cursor / other MCP clients
RepoPrimer does not try to remember every conversation and does not replace a knowledge base. It reads the project facts you choose to register and returns only the context an agent asks for.
MCP tools
The public MCP surface is intentionally limited to four read-only tools:
| Tool | Purpose |
|---|---|
list_projects |
Discover registered projects and their basic health. |
get_project_context |
Build a compact handoff from project docs and live Git state. |
search_project |
Search registered project Markdown and return bounded snippets. |
get_document |
Read one allowed project document with an optional size limit. |
Tools accept project identifiers, not arbitrary filesystem paths. RepoPrimer resolves and validates paths against the local registry before reading.
Quick start
Public package
The current release is a pre-release published under the alpha dist-tag, so
install it with @alpha rather than @latest.
Initialize or register a project explicitly:
npx -y @repoprimer/mcp@alpha init
npx -y @repoprimer/mcp@alpha context
Add the MCP server to Codex:
codex mcp add repoprimer -- npx -y @repoprimer/mcp@alpha serve
For long-lived MCP configuration, pin an exact tested version instead of a moving dist-tag.
Local development
RepoPrimer requires Node.js 22 or 24.
git clone https://github.com/repoprimer/repoprimer.git
cd repoprimer
npm install
npm test
node dist/cli.js doctor
node dist/cli.js serve
Project documents
RepoPrimer works with existing Markdown conventions. A minimal project can use:
docs/
|-- PROJECT_STATE.md
|-- TASKS.md
`-- DECISIONS.md
Projects are not required to adopt those exact filenames. A project config can
map existing files such as AGENTS.md, CLAUDE.md, ADRs, or a Memory Bank into
the document set that RepoPrimer may read.
Generated context is assembled on demand. It is not another source of truth.
Read-only and privacy boundaries
The MCP server:
- reads only registered projects and configured documents;
- reads Git metadata through bounded, non-mutating commands;
- does not write project files or alter the working tree;
- does not call model APIs, upload content, or collect telemetry;
- does not require a background database or cloud account.
The CLI has a separate, explicit setup boundary: commands such as init, add,
and remove may update RepoPrimer configuration or its user-level registry.
They are never invoked implicitly by an MCP read tool.
Scope
RepoPrimer v0.1 focuses on project discovery, live context, document retrieval, and text search. The following are intentionally out of scope:
- conversation capture and autonomous memory writes;
- embeddings, vector databases, or knowledge graphs;
- cloud sync, accounts, teams, or a hosted service;
- a web UI or IDE-specific extension;
- source-code indexing or autonomous repository analysis;
- model inference, API keys, or telemetry.
See Architecture, Decisions, and Tasks for the current design and implementation status.
FAQ
Why not just let the agent explore the repository itself?
It can, and RepoPrimer does not prevent that. The difference is the first
turn: exploration is re-run in every session by every agent, varies between
runs, and spends context-window tokens on rediscovery. RepoPrimer makes the
first turn deterministic and bounded — every agent starts from the same brief
within a fixed character budget, and deeper reads remain available through
get_document and search_project.
How does this relate to CLAUDE.md or AGENTS.md?
They are complementary. Those files hold durable instructions for one
repository, read by the agents that support them. RepoPrimer adds what a
static file cannot: live Git state, a multi-project registry, bounded
responses, and one context source shared by every MCP-compatible agent. A
project config can map CLAUDE.md or AGENTS.md into the document set
RepoPrimer serves.
Contributing
RepoPrimer is being prepared as an independent open-source project. Please read CONTRIBUTING.md before proposing a change and follow the Code of Conduct. Security issues should follow SECURITY.md, not the public issue tracker.
English is the canonical language for repository metadata and project governance. The complete Simplified Chinese README provides a secondary onboarding path, and issues or feedback in Chinese are welcome.
License
Licensed under the Apache License 2.0.
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。