# MCP reference Connect an AI assistant to a project through Datool's authenticated MCP server. ## Connect [#connect] Add your Datool instance's `/api/mcp` URL to an MCP client that supports Streamable HTTP and OAuth: ```text https://your-datool-host/api/mcp ``` The client discovers the authorization server and opens a browser. Sign in, select an organization and project, and approve the requested scopes. A grant is bound to that project; tool arguments cannot switch it to another project. For non-interactive agents, the same endpoint accepts an organization API key in `Authorization: Bearer ` together with `x-project-id`. The key needs explicit resource permissions; `reviews:write` permits submissions, `reviews:read` permits review reads, and `traces:read` permits inspecting evidence. Legacy ingestion keys cannot review. ## What an assistant can do [#what-an-assistant-can-do] Depending on granted scopes, an assistant can investigate traces and sessions, manage datasets and scorers, start evaluation runs, inspect metrics, and work with review sessions. Start with a concrete request, for example: > Find recent errored traces in this project, inspect a few representative failures, and summarize their recorded causes. The client should discover tools and their schemas from the server. Read-only investigation needs read scopes. Resource creation, updates, and evaluation execution need the relevant write scopes. ## Manage access [#manage-access] Open **Project settings → MCP connections** to inspect and revoke your connections. Revoking consent also revokes associated refresh tokens. Membership and authorization are rechecked for requests and refreshes. API-key and OAuth feedback in review sessions is **AI-labelled** and attributed to the authenticated principal. Optional agent/model metadata never establishes identity. AI completion is counted separately from human completion and does not update dataset ground truth. See [Reviews and Human Scores](/docs/guides/reviews). ## Claude Code setup [#claude-code-setup] Register the remote endpoint from the application directory: ```sh claude mcp add --transport http datool https://your-datool-host/api/mcp claude mcp get datool ``` Open Claude Code and use `/mcp` to select Datool and complete browser authentication. Choose the intended project and the scopes needed for your work. Verify the connection with a read-only request: “List the latest trace in this project and return its ID and status.” Confirm the project before allowing mutations. See the [official Claude Code MCP guide](https://code.claude.com/docs/en/mcp) for client-specific scope and authentication behavior. Other clients must support Streamable HTTP and Datool's advertised OAuth flow. ## Troubleshooting [#troubleshooting] * If sign-in is denied, confirm that your verified email domain is allowed and that you used a configured sign-in method. * If a tool returns a permission error, check the grant's scopes and your project access. * If you selected the wrong project, establish a new authorization for the intended project. * If a client cannot complete login, confirm that it supports the advertised OAuth flow and exact redirect registration. The [CLI](/docs/reference/cli) exposes the same operations through `datool agent tools` and `datool agent call`. ## Datool agent skills [#datool-agent-skills] The maintained [Datool skills repository](https://github.com/vinpac/datool-skills) provides task guidance for tracing, datasets, scorers, evaluations, and analytics. Follow its installation instructions for your agent. Skills guide tool use; they do not grant project access or replace MCP/CLI authentication.