repo-sensei

Connect tools

MCP in your editor

Connect Claude Code, VS Code or Cursor, with OAuth or an API key.

repo-sensei is an MCP server: your editor's coding agent can search the code, look up symbols, read files, trace features, walk the graph, check what a change reaches, and ask, all through the same index you use here.

Connect with OAuth (recommended)

Add the server to your editor; the first time the agent uses it, the editor opens a page here to sign in and approve access. Nothing to copy or keep.

  • Claude Code

    .mcp.json

    {
      "mcpServers": {
        "repo-sensei": {
          "type": "http",
          "url": "https://reposensei.com/mcp"
        }
      }
    }
    claude mcp add --transport http repo-sensei https://reposensei.com/mcp

    In Claude Code, run /mcp and choose repo-sensei to sign in.

  • VS Code

    .vscode/mcp.json

    {
      "servers": {
        "repo-sensei": {
          "type": "http",
          "url": "https://reposensei.com/mcp"
        }
      }
    }
    code --add-mcp '{"name":"repo-sensei","type":"http","url":"https://reposensei.com/mcp"}'

    Start the server from the MCP view and sign in when VS Code asks.

  • Cursor

    .cursor/mcp.json

    {
      "mcpServers": {
        "repo-sensei": {
          "url": "https://reposensei.com/mcp"
        }
      }
    }

    In Cursor's MCP settings, sign in to repo-sensei when it asks.

Your organization's Settings → Editors (MCP) has the same setup, and Test checks that this server answers an editor as it should.

Connect with an API key

To run the server on your machine instead (over stdio), install sensei-mcp (see the command line for how) and give it an API key:

  1. In Settings → API keys, create a key. Choose what it may do: read:repo (search, code, trace, graph and impact) and, if the agent should ask questions, ask. The key is shown once; keep it somewhere safe.
  2. Put it in the environment as SENSEI_API_KEY, and use the local setup:
  • Claude Code

    .mcp.json

    {
      "mcpServers": {
        "repo-sensei": {
          "type": "stdio",
          "command": "sensei-mcp",
          "env": {
            "SENSEI_URL": "https://reposensei.com",
            "SENSEI_API_KEY": "${SENSEI_API_KEY}"
          }
        }
      }
    }
    claude mcp add --env SENSEI_URL=https://reposensei.com --env SENSEI_API_KEY="$SENSEI_API_KEY" --transport stdio repo-sensei -- sensei-mcp

    Set SENSEI_API_KEY in the environment Claude Code starts from.

  • VS Code

    .vscode/mcp.json

    {
      "inputs": [
        {
          "type": "promptString",
          "id": "sensei-api-key",
          "description": "repo-sensei API key",
          "password": true
        }
      ],
      "servers": {
        "repo-sensei": {
          "type": "stdio",
          "command": "sensei-mcp",
          "env": {
            "SENSEI_URL": "https://reposensei.com",
            "SENSEI_API_KEY": "${input:sensei-api-key}"
          }
        }
      }
    }

    VS Code asks for the API key the first time the server starts, and keeps it.

  • Cursor

    .cursor/mcp.json

    {
      "mcpServers": {
        "repo-sensei": {
          "type": "stdio",
          "command": "sensei-mcp",
          "env": {
            "SENSEI_URL": "https://reposensei.com",
            "SENSEI_API_KEY": "${env:SENSEI_API_KEY}"
          }
        }
      }
    }

    Set SENSEI_API_KEY in the environment Cursor starts from.

Revoke a key at any time from the same page.

The tools

ToolWhat it does
sensei_list_reposThe repositories you can read, by slug.
sensei_searchExact, keyword and meaning search.
sensei_find_symbolWhere a symbol is defined.
sensei_read_codeA file's lines, numbered.
sensei_traceA feature, hop by hop across layers.
sensei_graph_neighborsWhat a symbol calls, and what calls it.
sensei_impactWhat a change to a symbol reaches.
sensei_askAn answer with its citations (spends the organization's budget).

Every tool only reads, and only the repositories you can read. What they return from the code is marked as the repository's content, to be read as data, not followed as instructions.