mcp-repo-context
A Weaviate-backed MCP server that gives Claude Code agents semantic search over your repository's PR review comments and source code.
README
mcp-repo-context
A Weaviate-backed MCP server that gives Claude Code agents semantic search over your repository's PR review comments and source code.
What it does
- Extracts PR review comments from GitHub and chunks your source code into functions/classes/types
- Embeds everything into Weaviate via the
text2vec-transformersmodule (runs locally, no API keys) - Serves 6 MCP tools that Claude Code agents can call during peer reviews and solution planning
MCP Tools
| Tool | Description |
|---|---|
search_similar_reviews |
Semantic search across past PR review comments |
get_review_patterns_for_file |
Find review feedback related to a specific file |
get_ticket_review_history |
All review comments for a specific ticket |
search_codebase |
Semantic search across source code chunks |
get_file_chunks |
Get all functions/classes in a file |
get_module_overview |
List all code in a module |
Prerequisites
- Weaviate running with the
text2vec-transformersmodule enabled and reachable atWEAVIATE_HOST:WEAVIATE_PORT(defaults:localhost:8080HTTP,localhost:50052gRPC). Any deployment works — a standalonedocker composestack, a hosted instance, etc. - Python 3 with
pip(needsweaviate-client>=4.0.0) - Node.js >= 18
- Claude Code CLI (
claude) - GitHub CLI (
gh), authenticated — only required if you extract fresh review data
Quick Start
The setup script is designed to be run from the root of the repo you want to index. It verifies Weaviate connectivity, optionally ingests data, builds the MCP server, and registers it with Claude Code.
Option A: Register against already-ingested Weaviate
cd /path/to/your-repo
/path/to/mcp-repo-context/setup.sh
Option B: Ingest pre-built JSONL first, then register
cd /path/to/your-repo
/path/to/mcp-repo-context/setup.sh --data-dir ./data
The --data-dir flag points at a directory containing review-comments.jsonl and/or codebase-chunks.jsonl (schema is defined by ingest.py).
Option C: Extract fresh data from GitHub, ingest, then register
cd /path/to/your-repo
REVIEW_AUTHORS=your-github-username /path/to/mcp-repo-context/setup.sh --extract
This runs extract-review-comments.py and extract-codebase.py against the current repo, writes JSONL into ./data/, ingests into Weaviate, and registers the MCP server.
Configuration
All scripts read configuration from environment variables.
Weaviate connection
| Variable | Default | Description |
|---|---|---|
WEAVIATE_HOST |
localhost |
Weaviate host |
WEAVIATE_PORT |
8080 |
Weaviate HTTP port |
WEAVIATE_GRPC_PORT |
50052 |
Weaviate gRPC port |
Extraction (when using --extract)
| Variable | Default | Description |
|---|---|---|
GITHUB_REPO |
Auto-detected via gh |
GitHub repo in owner/repo format |
REVIEW_AUTHORS |
All humans (bots excluded) | Comma-separated GitHub usernames to include |
TICKET_PATTERN |
[A-Z]+-\d+ |
Regex to extract ticket IDs from PR titles |
SOURCE_GLOBS |
**/*.ts, **/*.vue, **/*.js |
JSON array of [label, glob] pairs |
EXCLUDE_DIRS |
node_modules,dist,... |
Comma-separated directories to skip |
How it works
- Weaviate stores the data. Two collections are created by
ingest.py— one for review comments, one for codebase chunks. Vectors are generated at ingest time by Weaviate'stext2vec-transformersmodule, so no embeddings live in the JSONL. - Extraction is optional. The repo ships with
extract-review-comments.pyandextract-codebase.pyfor generating JSONL from scratch, but you can also hand-author JSONL files if you already have the data. - The MCP server runs on the host.
setup.shrunsnpm install && npm run buildagainstmcp-server/and registers the compiledbuild/index.jswith Claude Code viaclaude mcp add. No Docker is required for the server itself. - Queries go through a Python bridge. The Node MCP server spawns
query-review-knowledge.py/query-codebase.pyviaexecFile, passing JSON args and reading JSON results from stdout. This keeps the Weaviate client in Python while the MCP protocol is served over stdio from Node.
After setup
The setup script registers an MCP server named review-knowledge. To use its tools in Claude Code, allowlist the tool IDs in your Claude Code permissions:
mcp__review-knowledge__search_similar_reviews
mcp__review-knowledge__get_review_patterns_for_file
mcp__review-knowledge__get_ticket_review_history
mcp__review-knowledge__search_codebase
mcp__review-knowledge__get_file_chunks
mcp__review-knowledge__get_module_overview
Then restart Claude Code. Verify registration with claude mcp list.
License
MIT — 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。