Refd

What refd MCP server does

The refd MCP server exposes workspace intelligence from refd.ai to AI agents over the Model Context Protocol. A workspace represents one tracked brand along with its competitors, prompt set, monitoring runs, and enabled AI surfaces. Agents can retrieve the data used to understand how AI-generated answers discuss a brand.

The available intelligence covers visibility, brand and competitor mentions, citations, first-mention position, rankings, prompts, and answer evidence. This makes the refd MCP server relevant for research, search visibility analysis, reporting, and workflows that need to compare a brand's presence across monitored AI search surfaces.

How it works

The hosted remote connector is available at https://api.refd.ai/mcp. The default data:read scope is read-only. An optional data:write scope permits a limited setup lifecycle, can start one onboarding report, and can revoke the connection after explicit confirmation.

Refd stores workspace monitoring data generated from AI search surfaces including ChatGPT, Perplexity, Gemini, Google AI Mode, and Google AI Overviews. Monitoring runs score tracked entities for whether they were mentioned or cited and record first-mention position. AI Overview absence is represented as a valid result rather than treated as a failed request.

Setup and configuration

For the hosted refd MCP server, clients can connect through OAuth. The repository documents installation for Claude, Claude Code, ChatGPT, generic MCP clients, and headless agents. Cursor and VS Code have one-click installation links, and Claude Code can be configured with the remote HTTP URL. Headless agents use a personal access token created under Settings → Personal access tokens instead of OAuth.

Self-hosting requires a Cloudflare account, Workers, a D1 database, an R2 bucket, Queues, Workers AI, Browser Rendering, and a BrightData account. The deployment also needs configuration for the API origins, BrightData datasets and SERP zone, scheduled monitoring policy, and the listed Cloudflare bindings. Required secrets include JWT_SECRET, BRIGHTDATA_API_TOKEN, and ADMIN_EMAILS; EXA_API_KEY is optional for onboarding competitor search.

Tools and capabilities

The refd MCP server is suited to agents that need to:

• Read workspace-level AI search visibility data.
• Inspect tracked competitors and prompt sets.
• Review mentions, citations, rankings, and first-mention positions.
• Examine answer evidence from monitoring runs.
• Start the supported onboarding report when write scope is granted.
• Revoke a connection with explicit confirmation.

The hosted service uses a default read-only permission model, which limits mutation unless the optional write scope is intentionally enabled. Standard full runs can include up to 25 active prompts across three enabled surfaces with the default single sample, so results should be interpreted across completed runs rather than from one answer alone.