Nhost
io.github.nhost由 GitHub 账号 nhost 发布v0.4.0-beta2更新于 Nov 3, 2026
让用户通过AI助手处理自己的数据,且仅限于他们可见的内容。
AI 翻译由 AI 模型根据下方的原文描述翻译,可能有误。
原文描述
Let your users work with their own data through an AI assistant, limited to what they can see.
概览
Nhost 的本地 MCP 服务;无需登录;STDIO。
暂无 AI 概览:以下信息来自注册中心条目和服务本身。
安装
在 SourceWeft 中
- 打开 控制台中的 Nhost,将其添加到工作区。
- 为需要使用其工具的对话启用该服务。
Desktop only,通过 STDIO。 STDIO 服务会启动本地进程,因此需要 SourceWeft 桌面宿主。
其他 MCP 客户端
参照 仓库 中的启动说明。
README
Nhost MCP Service
The Nhost MCP service adds an agentic interface to any Nhost project. It exposes your project's GraphQL API as an MCP (Model Context Protocol) server, allowing AI assistants like Claude, Cursor, or any MCP-compatible client to query and mutate your data.
If you're building on Nhost, this service lets your users interact with their project data through AI assistants — no custom integration code required. Deploy it alongside your project and any MCP-compatible client can connect, authenticate on behalf of your users via your existing Nhost Auth setup, and work with your data within the permissions defined.
How It Works
The MCP server sits between your AI assistant and your Nhost project's GraphQL API. It provides three tools that the assistant can call:
get-schema- Introspects the GraphQL schema so the assistant understands your data model. Returns either a summary of available operations or the full SDL for specific queries/mutations.graphql-query- Executes read-only GraphQL queries against your project.graphql-mutation- Executes GraphQL mutations to create, update, or delete data.
On startup, the server fetches a schema summary from your GraphQL endpoint and includes it in the MCP instructions, giving the AI assistant immediate context about your data model without any additional configuration.
Authentication with OAuth2
The MCP service integrates with Nhost Auth as an OAuth2 authorization server, following the MCP Authorization specification. This means any MCP client that supports OAuth2 can authenticate with your Nhost project automatically.
The flow works as follows:
- The MCP client discovers authentication requirements by fetching
/.well-known/oauth-protected-resourcefrom the MCP server. - The client redirects the user to the Nhost Auth authorization endpoint for login (authorization code flow with PKCE).
- After authentication, the client exchanges the authorization code for a JWT access token.
- All subsequent MCP requests include the JWT as a Bearer token.
- The MCP server validates the JWT signature via JWKS, checks the issuer and expiration, then forwards the token on every GraphQL request.
- The GraphQL service enforces row-level and column-level permissions based on the JWT's claims, so the AI assistant can only access data the authenticated user is allowed to see.
IMPORTANT! CIMD needs to be enabled.
Enforcing a Specific Role
The --enforce-role flag restricts the MCP server to accept only tokens whose x-hasura-default-role claim matches the specified role. Requests with a different default role receive a 403 Forbidden response.
For example, setting --enforce-role=user_mcp means only tokens issued with user_mcp as the default role will be accepted. If you've configured permissions for that role with limited access (e.g., read-only on certain tables), the AI assistant will be restricted to those permissions.
This lets you:
- Create a dedicated role (e.g.,
user_mcp) with restricted permissions for AI access. - Allow AI assistants to read data but not modify it.
- Limit which tables or columns are visible to the assistant.
- Run multiple MCP instances with different role configurations for different use cases.
Configuration
All flags can be set via environment variables.
Deploying to Nhost
To deploy to Nhost Run you can use the one-click install link below:
And edit as needed. Make sure to set the correct subdomain and region in the MCP_AUTH_URL (or your custom domain if you using one), set the MCP_REALM to the URL you will use for your MCP service, change the MCP_INSTRUCTIONS as needed and change or remove MCP_ENFORCE_ROLE
NOTE: Check the releases page and make sure you are using the most recent version of the service.
Alternatively, add a run-mcp.toml to your project directory:
Then deploy your project as usual. The MCP server will be available at the published port and any MCP-compatible client can connect to it.
You can also try locally with:
Connecting an MCP Client
Point your MCP client to the server URL. For example, in Claude Desktop's configuration:
- Go to customize -> Connectors
- Click on
+and enter any name you want and the url where your MCP server resides, for instance https://mcp.acme.com - Click "Add"
- Once connected you will be redirected to your Oauth2's consent page to authorize the Oauth2 client.
来源:services/mcp/README.md,提交 75fff6a
工具
0版本历史
3- v0.4.0-beta2最新Nov 3, 2026
- v0.4.0-beta1Nov 1, 2026
- v0.3.0Sep 18, 2026
评分
已满足适用检查项的 50%
该评分衡量此收录条目的完整程度,依据注册中心条目、端点与公开来源计算得出。它不是对服务器本身的评价,也不表示服务器安全或当前可用:可连通性是带日期的检查结果,评分也不包含安装量或用户评价。
- 近 30 天内可连通25 分
请保持端点能够响应 MCP 握手;SourceWeft 每 30 天探测一次。
- 已声明认证方式10 分
请在注册中心条目中声明凭据请求头,或保持端点能够响应探测,以便判定其认证方式。
- 已列出工具15 分
请让服务器在无凭据时响应 tools/list,或发布可在沙箱中无凭据启动的软件包。
- 描述必需15 分
- 图标10 分
- 文档15 分
- 许可证10 分