obsidian-vault-mcp
Multi-tenant remote MCP server that exposes any GitHub-backed Obsidian vault to any MCP client via OAuth authentication.
README
obsidian-vault-mcp
Multi-tenant remote MCP server that exposes any GitHub-backed Obsidian vault to any MCP client (Claude iOS, Claude Code, Hermes, etc.). Users authenticate with their own GitHub account via OAuth and pick which repo to use as their vault — no per-user secrets are baked into the server.
How it works
- User adds the server URL to their MCP client.
- First request returns
401with aWWW-Authenticatepointing at/.well-known/oauth-protected-resource. - Client discovers the OAuth Authorization Server metadata, registers itself (Dynamic Client Registration, RFC 7591), and starts an Authorization Code flow with PKCE.
- The Lambda redirects the user to GitHub's OAuth consent screen.
- After GitHub callback, the Lambda renders a small page where the user picks which of their repos to use as the vault.
- The Lambda issues an MCP access + refresh token; the GitHub user token and repo selection are stored in DynamoDB keyed by the access token.
- MCP calls then operate against that repo using the user's GitHub token.
Token lifetimes: access 1h, refresh 30d, auth codes 10m, repo picker 30m.
Tools exposed
list_files— list a directory in the vaultread_file— read a markdown file (text) or image (MCP image content)write_file— create or update a file;encoding: "utf-8" | "base64"search— GitHub code search scoped to the user's selected vault repo
One-time setup: GitHub OAuth App
- github.com → Settings → Developer settings → OAuth Apps → New OAuth App.
- Homepage URL: anything (e.g. the repo URL).
- Authorization callback URL: leave as a placeholder for now; you'll set the
real value after the first
cdk deploy(it's printed as a stack output). - Save. Note the Client ID (public) and generate a new Client secret (treat as sensitive).
Deploy
npm install
cd lambda && npm install && cd ..
export GITHUB_OAUTH_CLIENT_ID=Iv1.xxxxxxxxxxxx
npx cdk bootstrap # first time per account/region
npx cdk deploy
The stack prints:
McpEndpoint— the URL clients connect toGitHubCallbackUrl— paste this back into the OAuth App's "Authorization callback URL" fieldGitHubOAuthClientSecretParam— the SSM Parameter Store name where the OAuth App's client secret must be storedSessionTableName— the DynamoDB table holding sessions, codes, and tokens
Set the OAuth App client secret in SSM Parameter Store:
aws ssm put-parameter \
--name /obsidian-vault-mcp/github-oauth-client-secret \
--value 'your_oauth_app_client_secret' \
--type SecureString
(Use --overwrite if you ever need to rotate it.)
Continuous deploy via GitHub Actions
The repo ships with .github/workflows/deploy.yml which runs cdk deploy on
every push to main, authenticating to AWS via GitHub OIDC (no long-lived
access keys in GitHub secrets).
One-time setup:
- Deploy the OIDC stack with local credentials:
If your account already has a GitHub Actions OIDC provider, pass the existing ARN:npx cdk deploy ObsidianVaultMcpOidcStacknpx cdk deploy ObsidianVaultMcpOidcStack \ -c existingGitHubOidcProviderArn=arn:aws:iam::<acct>:oidc-provider/token.actions.githubusercontent.com - Copy the
DeployRoleArnoutput. - In the GitHub repo's Settings → Secrets and variables → Actions:
- Secret
AWS_DEPLOY_ROLE_ARN= the role ARN from step 2 - Variable
AWS_REGION= the region you're deploying to - Variable
GITHUB_OAUTH_CLIENT_ID= your OAuth App's Client ID
- Secret
After that, every push to main triggers a deploy.
Client setup
In any MCP client that supports remote MCP + OAuth, add the McpEndpoint URL
(or its /mcp sub-path) as a server. The client handles the rest of the OAuth
dance and pops a browser for GitHub login and repo selection.
Repo permissions
The OAuth flow requests the repo scope so users can pick public or private
repos. Tokens are stored encrypted at rest in DynamoDB (AWS-managed KMS). Each
user only ever sees their own data; the server has no global vault.
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。