SourceRecall MCP Server
Enables AI coding assistants to search and retrieve code from self-hosted Git repositories using semantic and lexical search with optional reranking, while keeping repository data under the operator's control.
README
SourceRecall
SourceRecall is a self-hosted repository intelligence service for AI coding assistants. It indexes clean Git working trees into Qdrant, combines semantic and lexical retrieval, reranks candidates through a separate Jetson NLP service, and exposes the results through HTTP and Model Context Protocol (MCP).
It is designed for a small trusted-network deployment where repository data must remain under the operator's control.
Highlights
- Safe repository containment beneath one managed root
- Clean-commit indexing with file, chunk, model, and schema provenance
- Authoritative full re-indexing that removes obsolete Qdrant points
- Hybrid vector and ripgrep retrieval with chunk-aware reciprocal-rank fusion
- Optional cross-encoder reranking through Jetson NLP
- Thin Streamable HTTP MCP integration with structured tool responses
- Durable SQLite job and index status
- Validated YAML configuration and structured rotating-file logs
- Single-worker, serialized indexing model with explicit resource bounds
Architecture
Continue Agent
|
| Streamable HTTP MCP
v
SourceRecall MCP :8071
|
v
SourceRecall API :8070
| |
| +--> managed Git repositories
| + ripgrep lexical retrieval
|
+--> Qdrant :6333
+--> Jetson NLP :8091 (/v1/embeddings and /v1/rerank)
SourceRecall owns repository, path, commit, chunk, and retrieval semantics. Jetson NLP remains a stateless inference service that only knows text, vectors, and scores.
Repository layout
config/ Versioned configuration template
deploy/systemd/ API and MCP systemd units
diags/ Deployment validation utilities
docs/ Architecture, API, deployment, and operations guides
src/source_recall/ Installable Python package
tests/ Automated tests
API at a glance
curl http://source-recall:8070/live
curl -X POST http://source-recall:8070/index \
-H 'Content-Type: application/json' \
-d '{"repository":"episode-tracker"}'
curl -X POST http://source-recall:8070/search \
-H 'Content-Type: application/json' \
-d '{
"repository":"episode-tracker",
"query":"where are TV episode credits normalized?",
"limit":8
}'
When security.api_token is configured, add
Authorization: Bearer <token> to all non-health API requests.
Production paths
| Purpose | Path |
|---|---|
| Application | /opt/source-recall |
| Virtual environment | /opt/source-recall/.venv |
| Configuration | /etc/source-recall/source-recall.yaml |
| Managed repositories | /opt/source-recall/repositories |
| Durable state | /var/lib/source-recall/source-recall.db |
| Application logs | /var/log/source-recall/source-recall.log |
| API service | /etc/systemd/system/source-recall-api.service |
| MCP service | /etc/systemd/system/source-recall-mcp.service |
See Deployment for the complete installation procedure.
Documentation
- Architecture
- API reference
- Configuration
- Deployment
- Development
- MCP integration
- Operations
- Initial audit
Current scope
The initial release performs authoritative full indexing. Incremental Git indexing, webhook ingestion, syntax-aware chunking, symbol graphs, and reference analysis are planned extensions rather than claimed features. Indexing jobs are durable for status reporting but execute in one local process; this is not a distributed job queue.
The services do not terminate TLS, and the MCP endpoint does not implement client authentication. Keep them on a trusted private network or place them behind an authenticated reverse proxy. See Security.
License
SourceRecall is available under the MIT 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。