Obsidian MCP Assistant
Enables AI-powered chat and file operations on Obsidian vaults with local indexing, hybrid search, and MCP access.
README
Obsidian MCP Assistant
Desktop-only Obsidian plugin and local companion service for hybrid vault search, cited AI chat, attachment extraction, and read/write MCP access.
Everything indexed and retrieved stays on the machine. Cloud provider requests happen only after a provider, model, and extraction consent are configured. The companion binds to loopback only.
What ships
- Obsidian plugin — assistant pane, hybrid search modal, provider/model settings, extraction consent, MCP token management, rebuild and recovery controls.
- Local companion — the sole owner of index/state SQLite databases, watcher, work queue, retrieval, HTTP control API, and Streamable HTTP MCP endpoint.
- MCP — read, list, links, search, create, edit, move, delete, and reindex tools over either local stdio or authenticated loopback HTTP.
- Retrieval — deterministic BM25/FTS keyword search, optional embeddings, reciprocal-rank hybrid fusion, filters, source provenance, and snapshot-backed chat citations.
- Attachments — local Markdown/Canvas/text/source/office/PDF extraction plus opt-in provider handling for document, image, audio, and video content.
Requirements
- Obsidian desktop 1.11.4 or newer. Mobile is unsupported.
- Node.js 20 or newer for the companion.
- A supported macOS/Linux/Windows native-runtime environment for
better-sqlite3,sqlite-vec,@napi-rs/keyring, andsharp. - Optional provider credentials for OpenAI, Anthropic, and/or Google Gemini. Local keyword indexing does not require a provider.
Install from a release
-
Download the release assets:
obsidian-mcp-assistant-plugin-vX.Y.Z.zipobsidian-mcp-companion-X.Y.Z.tgz
-
Extract the plugin zip into your vault at:
<vault>/.obsidian/plugins/obsidian-mcp-assistant/ -
Enable Obsidian MCP Assistant in Obsidian’s Community plugins settings.
-
Install the companion from the tarball in a local directory:
npm install /path/to/obsidian-mcp-companion-X.Y.Z.tgz -
Register the vault. Save the one-time token printed by this command; it is shown only once.
npx obsidian-mcp init --vault /absolute/path/to/vault --id my-vault -
Paste that token into the plugin settings, then start the daemon:
npx obsidian-mcp serve --vault-id my-vault
The daemon listens only on 127.0.0.1. Keep it running while using the plugin or MCP clients.
Development install
npm ci
npm run build
npm run dev:install -- /absolute/path/to/vault
node packages/companion/dist/cli.js init --vault /absolute/path/to/vault --id my-vault
node packages/companion/dist/cli.js serve --vault-id my-vault
Run npm run test:native-package before relying on a local native build. It validates the packaged companion and confirms native dependencies are excluded from the Obsidian bundle.
Providers and consent
Provider credentials are validated before persistence and stored by the operating system credential manager, not in plugin data or vault files.
| Capability | Supported providers |
|---|---|
| Chat | OpenAI, Anthropic, Google |
| Embeddings | OpenAI, Google |
| Document/image/PDF extraction | OpenAI, Anthropic, Google |
| Audio extraction | OpenAI, Google |
| Video extraction |
Enable each extraction modality explicitly in the plugin settings. A credential alone does not permit uploads. Extraction failures remain visible as retryable jobs; unsupported/encrypted binaries remain metadata-only rather than being silently omitted.
MCP setup
Streamable HTTP
Create a separately revocable bearer token in plugin settings or with:
npx obsidian-mcp create-mcp-token --vault-id my-vault --name codex
Configure an MCP client to use:
http://127.0.0.1:<port>/mcp
Authorization: Bearer <mcp-token>
Never use the plugin-control token as an MCP bearer token.
Stdio
For local clients that support stdio, configure:
{
"command": "node",
"args": [
"/absolute/path/to/node_modules/@obsidian-mcp/companion/dist/cli.js",
"stdio",
"--vault-id",
"my-vault"
]
}
The stdio bridge requires the same vault daemon to be healthy; it never opens the vault databases itself.
Security model
- Loopback-only HTTP; non-loopback peers and unapproved origins are rejected.
- Plugin-control, MCP bearer, and private stdio-bridge credentials are distinct.
- MCP tokens are independently revocable. Revocation closes sessions bound to that token.
- Vault mutations require the current SHA-256 and are journaled with recovery records.
- Traversal, absolute paths, NUL bytes, symlinks, and configured excluded paths are rejected.
- Structured logs redact bodies, note content, tokens, keys, prompts, and provider credentials.
Operations
# Check native runtime support
npx obsidian-mcp doctor
# Persist a provider key; input is read from hidden stdin
npx obsidian-mcp set-key --vault-id my-vault --provider openai
# Remove a provider key
npx obsidian-mcp delete-key --vault-id my-vault --provider openai
# Revoke a single MCP token
npx obsidian-mcp revoke-mcp-token --vault-id my-vault --id <token-id>
Use the plugin settings for rebuilding the index, inspecting jobs, restoring journaled recovery items, rotating the plugin-control token, and session-only provider credentials.
Development commands
| Command | Purpose |
|---|---|
npm run typecheck |
Type-check all workspace packages. |
npx tsc -p tsconfig.json --noEmit |
Type-check integration tests. |
npm run lint |
Run strict ESLint checks. |
npm run format:check |
Verify Prettier formatting. |
npm run build |
Build companion, contracts, and plugin assets. |
npm test |
Run the default Vitest suite. |
npm run test:integration |
Run companion end-to-end tests. |
npm run test:native-package |
Fresh-install and native dependency verification. |
Release and versioning
Releases are GitHub assets, not npm publications. See VERSIONING.md for the synchronized-version and tag-driven release process, and CHANGELOG.md for release notes.
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。