生成引擎优化审计工具

Newtation MCP

Audit how AI systems perceive, describe, and cite your brand — from inside Claude.

Newtation MCP is a remote Model Context Protocol server for generative engine optimization (GEO). It gives Claude 14 tools that fetch live pages, grade AI answers against ground truth, and generate deployable fixes — plus a client tier that pipes deliverable and share-of-voice data from the Newtation dashboard.

The unique angle: the server runs inside the AI it's auditing, so features that competitors paywall (they burn API tokens per query) are free here.

Quick start

Connect from Claude Desktop by adding this to Settings → Developer → Edit Config:

{
  "mcpServers": {
    "newtation": {
      "command": "npx",
      "args": ["mcp-remote", "https://newtation-mcp.keshavparsai2.workers.dev/mcp"]
    }
  }
}

Restart Claude Desktop. Sign in with GitHub, then choose:

• Continue with free tools — all 11 auditing tools + 3 prompt workflows, unlimited
• I'm a client — sign in via dashboard — unlocks 3 additional tools that read from the Newtation dashboard (deliverables, live LLMrefs share-of-voice), scoped to your account

Works the same way in Claude Code, Cursor, and any other MCP client via mcp-remote.

The unique tools

These are the ones nobody else gives away free:

live_ai_perception_test (prompt workflow) Forces Claude to answer "What is {brand}?" from its own training memory before touching tools, then feeds that answer to score_ai_answer — a deterministic grader that fetches the brand's site + /about and returns an AI Knowledge Score (topic coverage + vocabulary alignment + entity fact-check), which competitors AI associates the brand with, and a knowledge lag report measuring how many product cycles behind AI's mental model is. If the AI doesn't know the brand at all, the tool returns a first-class "NOT RECOGNIZED — invisible to AI" verdict with entity-building actions.

hallucination_stress_test (prompt workflow) Makes Claude answer 6 risky prospect questions (pricing, founders, HQ, free tier, products, "still active?") from memory, then verify_brand_facts extracts every name, year, figure, and price and diffs it against homepage + /about + /pricing. Returns per-question verdicts, unverified claims, and an overall hallucination risk level. What competitors pay hundreds/month for.

generate_llms_txt Crawls the site with the same fetch pipeline as the auditors and emits a ready-to-deploy llms.txt (the llmstxt.org standard) with site name, description, Key Pages, and Optional sections built from the site's real internal links.

All tools

Free tier (11 tools + 2 prompt workflows)

Prompt workflows: live_ai_perception_test, hallucination_stress_test, run_seo_action_plan

Client tier (3 additional tools, dashboard-scoped)

Available after signing in with the email registered in the Newtation dashboard.

Authentication

Two verification layers:

1. GitHub OAuth 2.0 (PKCE) — proves ownership of a GitHub account
2. Tier choice — after step 1, you pick free tools or client tools. Choosing client redirects to app.newtationco.app where a Supabase magic-link login proves ownership of the dashboard email; that email is returned to the MCP server via a short-lived HMAC-signed token (5-min expiry, single-use) and drives all data scoping

Free-tier data never leaves Cloudflare; client-tier requests carry the verified email to the dashboard, which enforces per-client visibility server-side.

Example prompts

Run the live AI perception test for Notion (https://www.notion.com).

Stress-test what AI hallucinates about Stripe. URL: https://stripe.com

Generate a deployable llms.txt for my site at example.com.

Compare my site (example.com) against my top competitor (competitor.com) for the keyword "project management".

Self-hosting

git clone https://github.com/Elactrac/newtation-mcp.git
cd newtation-mcp
npm install

# Set secrets
npx wrangler secret put GITHUB_CLIENT_ID
npx wrangler secret put GITHUB_CLIENT_SECRET
npx wrangler secret put COOKIE_ENCRYPTION_KEY  # openssl rand -hex 32
npx wrangler secret put MCP_API_SECRET         # openssl rand -hex 32 — must match the dashboard

# Create KV namespace and paste the ID into wrangler.jsonc
npx wrangler kv namespace create "OAUTH_KV"

npm run deploy

cp .dev.vars.example .dev.vars   # fill in local OAuth credentials
npm run dev
# Test with: npx @modelcontextprotocol/inspector
# → http://localhost:8788/mcp

Project structure

newtation-mcp/
├── src/
│   ├── index.ts               MCP server (Durable Object) + tool registration
│   ├── tools.ts               All 14 tool implementations
│   ├── github-handler.ts      GitHub OAuth + tier-choice flow
│   ├── utils.ts               OAuth + HMAC helpers
│   └── workers-oauth-utils.ts CSRF, state, approval dialog
├── wrangler.jsonc             Cloudflare Workers config
├── smithery.yaml              Smithery registry manifest
└── worker-configuration.d.ts  Runtime types

Privacy & terms

• Privacy policy: https://newtationco.app/privacy
• All auditing tools are read-only — they fetch public URLs and generate reports
• OAuth tokens stored in Cloudflare KV, 7-day TTL
• Client-tier requests carry the dashboard-verified email to app.newtationco.app over HTTPS
• No conversation content is logged or persisted

Support

• Email: support@newtationco.app (48-hour response)
• Issues: https://github.com/Elactrac/newtation-mcp/issues
• Docs: https://github.com/Elactrac/newtation-mcp

License

MIT — free to use, fork, and modify.

Built by Newtation — AI presence management for modern brands.