AI Visibility MCP for Zed

In Zed, the raw answers six AI engines give are a question away from the assistant - ChatGPT, Perplexity, Google AI Overviews, Google AI Mode, Copilot and Gemini. Answer text and structured citations arrive as a tool result, in the same lightweight window as the docs and marketing copy those citations are about. One server, one narrow tool, no editor to switch to.

Connect AgentGEO to Zed

Zed configures MCP servers in its settings.json, under a context_servers key. Add an agentgeo entry:

{
  "context_servers": {
    "agentgeo": {
      "command": {
        "path": "npx",
        "args": ["-y", "agentgeo-mcp", "--key", "ag_live_..."]
      }
    }
  }
}

That's the whole install - the server is a zero-dependency npm package that speaks MCP over stdio on Node.js 18+, so there is nothing to compile and nothing to keep updated by hand. A live ag_live_ key comes with a plan. Want to prove the wiring before you spend anything? An ag_test_ key always returns clearly labelled demo records at zero credits. Save settings.json and the agentgeo tool becomes available to the assistant.

Every call returns one record per requested surface - the provider's answer, verbatim, with each source it cited as structured data:

{
  "surfaceKey": "perplexity",
  "status": "delivered",
  "answerText": "...",
  "sources": [
    { "title": "...", "url": "https://...", "position": 1 }
  ],
  "fetchedAt": "2026-07-06T09:12:04Z",
  "latencyMs": 8412,
  "providerRecordId": "..."
}

What to ask once it's connected

There are no arguments to memorize - say what you want in plain language. Name a market and the assistant passes country through (it defaults to US). language defaults to en but only reaches Google AI Overviews - the chatbot surfaces take a country and no language field.

• "Fetch what chatgpt and perplexity answer for our main category query, and tell me whether any cited domain is ours." The baseline visibility check, start to finish, without opening a browser.
• "Pull all six surfaces for the query this landing page targets and append the citation lists to geo-log.md." The evidence lands in the repo it describes.
• "Compare Perplexity's cited sources against the pages under docs/ and tell me which three are closest to earning a citation." Then edit them in the same window.
• "Does google_ai_overview cite our pricing page for our product-plus-pricing query? If not, rewrite the opening paragraph so it answers the question directly."
• "Diff gemini against copilot for this query and list the competitor domains both of them cite."
• After the copy change ships, re-fetch the same query and diff answerText and sources[] against last week's record - the check and the fix live in one file tree.

What the MCP server gives you

Deliberately small. A native, fast editor deserves a data layer that doesn't drag an analytics platform in behind it.

• One narrow tool - fetch_raw_answers, and nothing else. No tool sprawl in the assistant's context, one contract to learn.
• Six engines behind one contract - ChatGPT, Perplexity, Gemini, Google AI Overviews, Google AI Mode and Copilot share an identical record shape, so widening a check is one more key in surfaces.
• Structured sources[] - title, URL and position for every citation, ready to diff, count or grep. Nothing regexed out of prose.
• Raw records only - no rankings, no sentiment, no visibility score. The judgment stays with the assistant that can see your actual pages.
• Managed-scraper engine, maintained for you - the server posts to https://api.agentgeo.org/v1/fetches; keeping six collection paths healthy is AgentGEO's problem, not a headless browser in your project directory.
• Usage billing with a spend cap - one credit per delivered record, failed records cost zero, never per-seat.

What it doesn't do

AgentGEO is a GEO data layer, not a platform, and the narrowness is deliberate. Conclusions drawn outside your repo are worth less than raw records analyzed inside it.

• No dashboards, no scores, no recommendations engine. Records arrive as a tool result; the report is whatever the assistant writes into a file you can commit.
• Not instant. Live surfaces are slow. Requests wait up to 180 seconds, and an AI Overview SERP round-trip alone runs 40-90s - a call spanning several surfaces will sit on the slow ones. Let it finish.
• Some fetches finish async. A slow chatbot scrape can exceed the sync budget and come back failed with a providerFields.snapshot_id. Retry with that id and the same single surface to collect the finished answer without paying for a re-scrape (not valid for google_ai_overview).

Prefer the REST API?

The MCP server is a thin wrapper over one endpoint: POST https://api.agentgeo.org/v1/fetches. The Python and curl pages show the same request shape - handy when a check you first asked for in the editor graduates into a script that runs on a schedule.

Who builds on this

Developers who chose Zed because it starts instantly and stays out of the way, and who keep their docs and marketing copy in the same repo as the product. If your GEO work is really a stack of markdown files, wiring the six engines into the editor beats logging into someone else's dashboard to be told which file to open. See where you stand first with the free audit if you like.