AI Visibility MCP for Codex

In Codex, the raw answers the six engines actually give are one ask away: describe the check and fetch_raw_answers returns the answer text plus every cited source as a tool result. One MCP server, one deliberately narrow tool — and once it's added, a six-engine citation sweep is just another repeatable agent task, not a manual chore.

Connect AgentGEO to Codex

codex mcp add agentgeo -- npx -y agentgeo-mcp --key ag_live_...

That is the whole install. codex mcp add registers the server, npx pulls agentgeo-mcp from npm (stdio transport, zero npm dependencies, Node.js 18+), and the agentgeo tool is available to Codex's agent runs. Create your ag_live_ key in the quickstart — free tier, no credit card. Want to verify the wiring before fetching anything real? An ag_test_ key always returns clearly labelled demo records at zero credits.

When Codex calls fetch_raw_answers, each engine comes back as its own per-surface record — the raw answer, its sources in cited order, timing and provider metadata (the full record also carries the raw providerFields):

{
  "surfaceKey": "perplexity",
  "status": "delivered",
  "answerText": "...",
  "sources": [
    { "title": "...", "url": "https://...", "position": 1 }
  ],
  "fetchedAt": "2026-08-18T09:12:04Z",
  "latencyMs": 8412,
  "providerRecordId": "..."
}

What to ask once it's connected

No arguments to memorize — describe the task and Codex shapes the call (country defaults to US, language to en; say "in Japan, in Japanese" to change it). Asks that make good repeatable agent tasks:

• "Fetch what ChatGPT and Gemini answer for 'best CI/CD for monorepos' and list which domains each cites."
• "Run our top three category queries across all six surfaces and report which ones cite ourdomain.com, with positions."
• "Pull Perplexity's answer for our main query, compare sources[] against geo-baseline.json, and print what entered or dropped out."
• "Diff google_ai_overview against google_ai_mode for 'best feature flag tools' and summarize the difference in cited sources."
• "Fetch Copilot's answer for 'alternatives to our product' — are we in the answer text, or only in the sources?"
• "Sweep all six engines for our five money queries and write a citation table to reports/geo-$(date).md." Wire that into your task runner and the check repeats itself.

What the MCP server gives you

Deliberately thin — an answer-access layer for your agent, not an analytics product bolted onto your CLI.

• One narrow tool — fetch_raw_answers posts to AgentGEO's API and returns the run envelope verbatim. Nothing else to learn, nothing else to break.
• Six engines behind one contract — ChatGPT, Perplexity, Gemini, Google AI Overviews, Google AI Mode and Copilot with an identical record shape, so adding an engine to a task is one more word in the ask.
• Structured `sources[]` — every citation with its title, URL and position, ready for the agent to diff, count or grep against your domain. No links regexed out of prose.
• Raw records only — no rankings, sentiment or visibility scores. The conclusions come from Codex, with your project in context.
• Managed-scraper engine, maintained for you — under the hood the server talks to POST https://api.agentgeo.org/v1/fetches; keeping six collection paths healthy is AgentGEO's job, not a headless browser in your task runner.
• Usage-based billing with a spend cap — one credit per delivered record, failed records cost zero, free tier with no credit card, never per-seat.

What it doesn't do

The narrowness is the point. Codex holds your project and your intent; conclusions computed outside that context would be worth less than raw records analyzed inside it — that is the whole GEO data layer argument. So, honestly:

• No dashboards. Records arrive as tool results; the report is whatever Codex writes from them, in your repo or your task output.
• No visibility scores or sentiment. Raw answerText and sources[] only — the judgment layer is deliberately left to your agent.
• No recommendations engine. Codex reasons over your actual pages, which beats generic advice every time.
• Not instant. Requests wait up to 180 seconds — live surfaces are slow. An AI Overview SERP round-trip runs 40–90s, and some chatbot scrapes exceed the sync budget entirely: those records come back failed with a providerFields.snapshot_id, and a follow-up call with that id and the same single surface redeems the finished answer without paying for a re-scrape (not valid for google_ai_overview). Budget for it when the task is unattended.

Prefer the REST API?

The MCP server is a thin wrapper over the same endpoint your scripts can hit directly: POST https://api.agentgeo.org/v1/fetches. See the Python and curl pages for the request shape — handy when an agent task graduates into a plain cron job with the identical contract.

Who builds on this

Developers who drive Codex for agentic tasks and want answer-engine checks to be one more thing the agent just does. Add the server once, and asking Codex to sweep the six engines for your category — and diff the citations against last week — becomes a repeatable task rather than a manual afternoon. Try it first in the no-signup playground, or see how the approaches compare.