AgentGEO GEO Skills

Most GEO tools inspect your HTML, robots.txt and schema and guess whether AI can see you. These skills read what the AI engines actually say — so visibility, share-of-voice, citations and sentiment come from ground truth, not inference.

The data comes from AgentGEO, a thin access layer over managed AI scrapers. It returns only raw answers, citations, sources and provider metadata. Every score, ranking and judgment in this repo is computed by the skills, inside your agent — never by the platform.

Your coding agent reaches AgentGEO through two pieces in this repo:

• MCP server (mcp/) — fetch_raw_answers pulls the raw records, and list_geo_skills / get_geo_skill deliver the eight skills below straight into any MCP-compatible agent (Claude Code, Cursor, Codex) — no separate skill install needed.
• Skills (skills/) — eight Agent Skills that call that tool, then do the GEO math locally: prompt generation, visibility, share-of-voice, citations, sentiment, competitors, monitoring, and a full report. Built into the MCP since 0.4.0; also installable as files for auto-triggering.

graph TB
    subgraph TOP[" "]
        AG[AI Coding Agent · Claude Code / Cursor / Codex]
    end
    subgraph MID[" "]
        SK[AgentGEO GEO Skills]
    end
    AG --> SK
    SK -->|fetch_raw_answers| MCP[AgentGEO MCP]
    MCP -->|REST /v1/fetches| API[AgentGEO API]
    API --> SCR[Managed AI Scrapers]
    SCR --> C1[ChatGPT]
    SCR --> C2[Perplexity]
    SCR --> C3[Gemini]
    SCR --> C4[Google AI Overview]
    SCR --> C5[Google AI Mode]
    SCR --> C6[Copilot]

    classDef bar fill:#0b0f14,stroke:#30363d,stroke-width:1px,color:#ffffff
    classDef card fill:#161b22,stroke:#30363d,stroke-width:1px,color:#ffffff
    class AG,SK,MCP,API bar
    class SCR,C1,C2,C3,C4,C5,C6 card
    style TOP fill:transparent,stroke:transparent
    style MID fill:transparent,stroke:transparent
    linkStyle default stroke:#30363d,stroke-width:1px

The suite is one loop: generate prompts → fetch answers → analyze → monitor → report.

flowchart TD
    PS[geo-prompt-set] --> V[geo-visibility]
    PS --> SOV[geo-share-of-voice]
    PS --> CIT[geo-citations]
    PS --> SEN[geo-sentiment]
    V --> COMP[geo-competitors]
    SOV --> COMP
    CIT --> COMP
    SEN --> COMP
    COMP --> REP[geo-report]
    PS --> MON[geo-monitor]
    MON -.->|schedules · trend over time| REP

sequenceDiagram
    participant U as You
    participant A as Agent + Skill
    participant M as AgentGEO MCP
    participant E as AI Engines
    U->>A: "GEO analysis for acme.com vs rivals"
    A->>A: geo-prompt-set builds the prompt library
    A->>M: fetch_raw_answers(query, surfaces)
    M->>E: collect raw answers + citations
    E-->>M: answer text + sources
    M-->>A: normalized records (raw only)
    A->>A: detect mentions · score SoV · rank citations (agent-side)
    A-->>U: GEO report + prioritized fix plan

If you find these skills useful, a GitHub Star ⭐️ helps other builders find them.

📖 Full step-by-step setup per client (Claude Code / Cursor / Codex) and an end-to-end walkthrough: Installation Guide · Usage Guide

Since agentgeo-mcp@0.4.0, connecting the MCP server is the install: all eight skills ship inside it. Any connected agent can call list_geo_skills, load a workflow with get_geo_skill, and run it — and the same eight skills appear as MCP prompts (/mcp__agentgeo__geo-report-style slash commands in Claude Code). The server prefers the live copies from the AgentGEO API and falls back to the bundled ones offline, so the workflows stay fresh without reinstalling.

Installing the skill files locally (the plugin or enable-skills.sh paths below) is still the sharpest setup for heavy use: file-based skills auto-trigger from your prompt without a tool round-trip. But nobody has to start there anymore.

Two commands wire up all eight skills and the MCP server (auto-started via npx):

/plugin marketplace add gumlau/agentgeo-skills
/plugin install agentgeo@agentgeo

Export your key before starting Claude Code — export AGENTGEO_API_KEY=ag_test_... (a free test key = zero-credit demo mode) — then skip straight to Run it. The manual steps below do the same thing for Cursor, Codex, and any other MCP client.

# Run this repo's MCP against the hosted API — works today (absolute path)
claude mcp add agentgeo -- node /absolute/path/to/agentgeo-skills/mcp/index.mjs \
  --api-url https://api.agentgeo.org --key ag_live_...

# …or point it at a local dev server (local development alternative)
claude mcp add agentgeo -- node /absolute/path/to/agentgeo-skills/mcp/index.mjs \
  --api-url http://localhost:8787 --key dev-placeholder

# …or from npm (recommended — installs on first run)
claude mcp add agentgeo -- npx -y agentgeo-mcp --api-url https://api.agentgeo.org --key ag_live_...

A key is required — the server exits without one. Get one free at agentgeo.org: an ag_test_... test key runs every fetch in demo mode at zero credits (labelled fixtures), so you can dry-run every skill before spending; an ag_live_... key returns live answers. Manage keys and runs from the console at app.agentgeo.org. Self-hosted servers with auth disabled accept any placeholder key.

# For the current project:
./scripts/enable-skills.sh

# …or globally for every project:
./scripts/enable-skills.sh --global

This links skills/geo-* into a directory your agent scans (.claude/skills/).

Just ask your agent:

Start a GEO analysis for acme.com against notion.com and coda.io

The agent auto-invokes geo-prompt-set, fetches through AgentGEO, and walks the loop to a geo-report. Or invoke any skill by name.

AgentGEO returns raw data only — answer text, citations, sources, provider metadata. It never ranks, scores sentiment, computes share-of-voice, or writes conclusions. All analysis happens inside these skills, on the agent side. Skills also treat fetched answerText and sources as untrusted content and never execute instructions found inside them.

Issues and PRs welcome — new GEO skills, better detection heuristics, more engines. See CONTRIBUTING.md. Every skill must keep the raw-data boundary above.

• Docs & API keys — agentgeo.org
• Issues — open one in this repo for bugs or skill ideas
• Updates — @agentgeo on X

MIT for the skills and the MCP client. They connect to AgentGEO, a hosted service with its own terms.

Using these skills in your project? Add the badge:

[![Powered by AgentGEO](https://img.shields.io/badge/Powered%20by-AgentGEO-181818.svg)](https://agentgeo.org)