Skip to main content

Connect external MCP tools

RAGFlow agents can use an external MCP server alongside their existing tools. This is different from running RAGFlow's MCP server, which exposes your RAGFlow resources to other clients.

Add web research with Parallel

Parallel Search MCP provides web_search for finding web sources and web_fetch for reading known public URLs. Its anonymous endpoint does not require a Parallel account or API key. Free access is rate limited.

  1. Save the example configuration as a JSON file:

    {
      "mcpServers": {
        "parallel-search": {
          "type": "streamable-http",
          "url": "https://search.parallel.ai/mcp",
          "headers": {
            "User-Agent": "ragflow"
          }
        }
      }
    }

  2. Open Settings → MCP and import the file. RAGFlow connects to the server and discovers its tools. Use Streamable HTTP, not SSE, if entering the connection manually. Leave the Parallel authorization token empty.

  3. In an Agent component, add the MCP tools from parallel-search and enable web_search and web_fetch as needed. Importing the server does not replace an existing search tool or make Parallel the default.

  4. Try a public-web question such as: “Find RAGFlow's current release notes and summarize the latest changes with source links.” The agent can search, then fetch a relevant result. No private dataset is needed for this example.

The User-Agent identifies RAGFlow so Parallel can measure aggregate usage by project. Keep it project-wide; do not add a user, installation or device identifier. Imported headers are used during discovery and subsequent tool calls and are retained when exporting the server configuration.

When enabled, the agent may call these tools during its normal workflow. Supplied search queries, requested URLs, objectives/context and request metadata go to Parallel. Review its privacy policy before sending sensitive content. Anonymous Parallel access does not remove authentication required by your RAGFlow installation.

Check a connection

Confirm that tool discovery returns web_search and web_fetch. A successful connection alone does not prove that the agent selected a tool; inspect the agent run's tool calls and returned source links.

For connection errors, check the transport, outbound access to search.parallel.ai, and the configured headers. The free example deliberately omits authentication headers. If you configured an authenticated connection, investigate its credentials instead of silently removing them. Respect rate-limit responses and avoid repeated immediate retries.

For other MCP servers, use their documented URL, transport and authentication. Custom headers must be an object with string names and string values. Never commit real credentials in an exported configuration.