# openDB — Local Memory DB + MCP for Agents

> openDB gives agents a local SQLite memory database with MCP tools: index a workspace, search files, then run `opendb serve-mcp` to expose 12 tools.

## Install

Merge the JSON below into your `.mcp.json`:

## Quick Use

```bash
pip install open-db[cli]
opendb index ./my_workspace
opendb serve-mcp
# example stdio server args:
opendb serve-mcp --workspace ./my_workspace
```

## Intro

openDB gives agents a local SQLite memory database with MCP tools: index a workspace, search files, then run `opendb serve-mcp` to expose 12 tools.

**Best for:** local-first agent memory + workspace search without external vector DBs

**Works with:** Python 3.11+, MCP clients, Claude Code/Cursor, Agents SDKs (examples in README)

**Setup time:** 5-12 minutes

### Key facts (verified)

- GitHub: 87 stars · 19 forks · pushed 2026-05-12.
- License: MIT · owner avatar + repo URL verified via GitHub API.
- README-verified entrypoint: `pip install open-db[cli]`.

## Main

- Keep the setup minimal: follow the README’s 3-line path (install → index → serve-mcp) before tuning workspaces and tool permissions.

- Run indexing per workspace so the agent can switch contexts cleanly; the README shows passing `--workspace` when serving MCP.

- Use this when you want local retrieval without embedding APIs: README emphasizes SQLite FTS5 + engineered ranking and includes benchmark numbers.

### Source-backed notes

- README opens with a 3-command setup: `pip install open-db[cli]`, `opendb index`, and `opendb serve-mcp`.
- README states the MCP server exposes 12 tools and provides config examples for clients (Claude Code/Cursor and multiple agent SDKs).
- README reports LongMemEval numbers (93.6% E2E accuracy; 500 questions) and highlights a SQLite-only retrieval stack.

### FAQ

- **Do I need a vector database?**: No — README emphasizes SQLite-only retrieval (FTS5) with engineered ranking.
- **How do I expose tools to my agent?**: Run `opendb serve-mcp` and wire it as an stdio MCP server (README shows multiple client examples).
- **Can I index multiple workspaces?**: Yes — README describes indexing and serving with an explicit `--workspace` path.

## Source & Thanks

> Source: https://github.com/wuwangzhang1216/openDB
> License: MIT
> GitHub stars: 87 · forks: 19

---

<!-- ZH -->

## Quick Use

```bash
pip install open-db[cli]
opendb index ./my_workspace
opendb serve-mcp
# example stdio server args:
opendb serve-mcp --workspace ./my_workspace
```

## Intro

openDB 给 agent 提供本地 SQLite 记忆数据库 + MCP 工具：`pip install` 后索引工作区、全文检索，并通过 `opendb serve-mcp` 以 stdio 接入；README 标注 12 个工具与 LongMemEval 93.6%。

**Best for:** 本地优先的 agent 记忆/检索，不想引入外部向量库的人

**Works with:** Python 3.11+；MCP 客户端；Claude Code/Cursor；Agents SDK 示例（README）

**Setup time:** 5-12 minutes

### Key facts (verified)

- GitHub：87 stars · 19 forks；最近更新 2026-05-12。
- 许可证：MIT；作者头像与仓库链接均已通过 GitHub API 复核。
- README 中核对过的入口命令：`pip install open-db[cli]`。

## Main

- 先按 README 的三步走（install → index → serve-mcp）跑通最小闭环，再考虑 workspace/权限等高级配置。

- 按 workspace 维度做索引，让 agent 能清晰切换上下文；README 展示了在 MCP server 中传 `--workspace`。

- 如果你想要“无 embedding API”的本地检索，这是合适选择：README 强调 SQLite FTS5 的方案，并给出基准数据。

### Source-backed notes

- README 一开始就给出 3 行安装路径：`pip install open-db[cli]`、`opendb index`、`opendb serve-mcp`。
- README 写明 MCP server 暴露 12 个工具，并提供 Claude Code/Cursor 与多种 Agents SDK 的接入示例。
- README 报告 LongMemEval 数据（93.6% E2E，500 题），并强调检索栈只用 SQLite。

### FAQ

- **需要向量数据库吗？**：不需要。README 强调用 SQLite FTS5 做检索，并给出性能/效果数据。
- **怎么把工具暴露给 agent？**：运行 `opendb serve-mcp` 并按 README 示例以 stdio MCP server 接入即可。
- **能管理多个 workspace 吗？**：可以。README 展示了索引与启动时通过 `--workspace` 指定路径的方式。

## Source & Thanks

> Source: https://github.com/wuwangzhang1216/openDB
> License: MIT
> GitHub stars: 87 · forks: 19


---
Source: https://tokrepo.com/en/workflows/opendb-local-memory-db-mcp-for-agents
Author: MCP Hub