MCP 服务器

optifeed-radar

Open-source CLI + MCP server that asks real AI engines (ChatGPT, Claude, Gemini, Perplexity) real buyer questions and scores whether a brand or its products get recommended (AI visibility / GEO / AEO). Runs locally with your own API keys.

README

Optifeed Radar

Open-source AI visibility checker. Now on npm - run it with npx optifeed-radar.

Is your brand recommended when buyers ask AI? Optifeed Radar checks whether ChatGPT, Perplexity, Gemini and Claude actually recommend you, and tells you where you stand against competitors. It runs locally, uses your own API keys, and has no Optifeed-hosted backend.

It is built for two kinds of AI agents at once: it measures how AI agents see and recommend you, and it can be run by your own AI agents (CLI, JSON, and an MCP server). People also call this AI visibility, generative engine optimization (GEO), answer engine optimization (AEO), or AI-SEO.

60-second setup

No install needed - npx fetches and runs it. The zero-key audit runs end to end with no API keys and no AI calls:

npx optifeed-radar audit yourbrand.com

It checks AI-crawler access (robots.txt), llms.txt, schema.org structured data, meta basics, and your sitemap, then prints a 0-100 AI-readiness score.

The check pipeline runs once you set at least one engine API key. Put it in a .env file in the directory you run from, or export it:

echo "OPENAI_API_KEY=sk-..." > .env      # any one engine key gets you started
npx optifeed-radar check yourbrand.com

The CLI loads .env from the directory you run it in, so there is no shell setup step. Exporting the keys works too (export OPENAI_API_KEY=...), and an exported key always wins over the same key in .env. config shows which keys were found and which file they came from, never the values.

It discovers your brand, generates a buyer-prompt pack, asks the engines, and scores recommendation, position, and share of voice into one AI Visibility Score. The score reads only the unbranded buyer questions (did the AI surface you unprompted); questions that name your brand are reported separately as reputation. All four engines are verified live against their production APIs (2026-07-20).

Working from a clone instead? Run npx tsx src/cli/index.ts <command> so flags reach the CLI unchanged, or use the npm run dev script with -- before the arguments (npm run dev -- check yourbrand.com --report out.html).

See it in action

Run a full visibility check from the terminal, from brand discovery and buyer prompt generation through live engine queries and scoring.

<a href="https://github.com/optifeed/optifeed-radar/blob/main/docs/assets/optifeed-radar-cli.mp4">Watch the 15-second CLI demo</a>

What it does

Optifeed Radar asks real AI engines real buyer questions and measures whether your brand gets recommended - not whether you rank in a search index, but whether the answer an AI gives a buyer names you. Grounded engines (which cite web sources) are reported separately from parametric ones (which answer from model weights alone), because they behave differently. An engine counts as grounded only for the answers where it actually searched: asking for grounded mode is a request a model can decline, so the report says when an engine searched on only some of its answers. METHODOLOGY.md has the formula.

The questions match what you sell. If you make your own products, buyers are asked what to buy and you are measured against rival makers. If you are a shop selling other companies' products, buyers are asked where to buy and you are measured against rival shops - product questions get answered with manufacturers, so scoring a shop on them reports a zero that says nothing about the shop. The tool works this out from your site and stores it as businessType in profile.json; edit it if it guessed wrong.

One level down, shopping does the same thing for individual products you name (beta). Each product gets its own 0-100 visibility score, and the report is ordered by what the engines did: any product they answered about but never recommended leads, since that is the finding worth reading, then the rest by visibility, and last anything the run could not measure at all. The order you list your products in carries no ranking meaning; it only breaks ties between identical scores. Each product is checked twice over - category buying questions that never name it, and questions that do - and when a product is absent the report leads with the rival products the engines named instead, which is the more useful half of a zero. Because every product is asked its own questions, the scores say how decisively each one wins its own shelf, not that one product beats another. You name the products; nothing is imported or crawled.

Use it from your AI agents (MCP)

The optifeed-mcp server exposes the same capability to AI agents. It runs over stdio, and npx fetches it on demand - no clone or build needed.

Claude Desktop (claude_desktop_config.json). The fastest way to open it is Settings -> Developer -> Edit Config, which creates the file if it does not exist yet. On disk it lives at:

macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
Windows: %APPDATA%\Claude\claude_desktop_config.json

{
  "mcpServers": {
    "optifeed-radar": {
      "command": "npx",
      "args": ["-y", "--package=optifeed-radar", "optifeed-mcp"],
      "env": {
        "OPENAI_API_KEY": "sk-..."
      }
    }
  }
}

Claude Desktop reads that file at startup, so quit and reopen it after editing.

Claude Code (.mcp.json in your project):

{
  "mcpServers": {
    "optifeed-radar": {
      "command": "npx",
      "args": ["-y", "--package=optifeed-radar", "optifeed-mcp"]
    }
  }
}

Cursor (.cursor/mcp.json) and Windsurf (mcp_config.json) use the same shape:

{
  "mcpServers": {
    "optifeed-radar": {
      "command": "npx",
      "args": ["-y", "--package=optifeed-radar", "optifeed-mcp"]
    }
  }
}

Working from a clone instead? Build first (npm install && npm run build), then run the server with node pointed at the built entrypoint - replace /path/to/optifeed-radar with your clone path:

{
  "mcpServers": {
    "optifeed-radar": {
      "command": "node",
      "args": ["/path/to/optifeed-radar/dist/mcp/index.js"]
    }
  }
}

Example prompts

Once it is connected, ask your AI agent in plain language. These map onto the five tools and the arguments they accept:

"Run a free AI-readiness audit on yourbrand.com." -> audit_store, no keys, no cost.
"What buyer questions should yourbrand.com be visible for?" -> generate_buyer_queries, so you can review the pack before paying for a run.
"Check yourbrand.com's AI visibility, quick mode, cap it at 20 cents." -> check_visibility with quick and max_cost.
"Check yourbrand.com on OpenAI and Perplexity only." -> check_visibility with engines.
"Check whether AI recommends my products: Aria 2, Presto X, Brew Mini, in that order, for yourbrand.com." -> shopping_check. Input order is not a ranking; it only breaks ties between products with identical scores.
"Check the Aria 2, a quiet home espresso machine, and the Presto X, a fast dual-boiler, on yourbrand.com. Cap it at one dollar." -> shopping_check with a descriptor per product and max_cost. Saying what each product IS is what rescues an opaque model name: without it the questions are guessed from the store category.
"What changed since the last visibility run on yourbrand.com?" -> get_snapshot_diff, free, and it needs two saved runs before it can compare.

Start with the audit prompt: it needs no keys, so it confirms the server is wired up before anything spends API credit. check_visibility runs non-interactively (no confirmation prompt over MCP), so the max_cost cap is what bounds a run your AI agent starts - it defaults to $0.50. shopping_check is bigger, so its default cap scales with the list, at $0.20 per product you name.

Tools and cost

Surface	Name	What it does	Cost
CLI	`audit`	Zero-key AI-readiness check (robots, llms.txt, schema, sitemap)	Free, no AI calls
CLI	`check`	Full pipeline: buyer prompts, engines, AI Visibility Score	BYO keys
CLI	`shopping`	Products you name: your ranking vs AI's, and the rival shelf	BYO keys
CLI	`diff`	What changed between your last two runs	Free (reads a saved snapshot)
CLI	`sources`	Domains the AI cited, and your share of voice	Free (reads a saved snapshot)
CLI	`queries`	Show or export your buyer-prompt pack	Free
CLI	`config`	Which engine keys are set, where state is stored	Free
MCP	`check_visibility`	Run a visibility check for a domain	BYO keys
MCP	`audit_store`	Run the zero-key readiness audit	Free
MCP	`generate_buyer_queries`	Produce the buyer-prompt pack	BYO keys (usually under $0.05)
MCP	`shopping_check`	Same product check, for your AI agents	BYO keys
MCP	`get_snapshot_diff`	Compare two saved runs	Free

Cost transparency: audit queries no AI engines and costs nothing. check spends your own API credit. Measured on real runs (2026-07-20, --quick = 8 buyer prompts):

run	measured cost
`audit`	free
`check --quick`, one engine	about $0.09
`check --quick`, all four	$0.41 to $0.46
`check --quick --grounded`, four	$0.85 to $1.09

Your cost varies with engine, prompt-pack size, and provider pricing. Grounded runs cost roughly 3x parametric ones, because web search is billed on top of tokens: Google charges per search query, and one answer can trigger several. The grounded range spans three real runs: two finished clean at about $0.85, and one that was reined in by its own cap spent $1.09, so treat the top of the range as the planning number.

shopping was measured on 2026-07-23: two products across all four engines in --grounded mode, 32 answers, $0.70 for the run - about $0.35 per product. That is the expensive corner (grounded runs cost roughly 3x parametric ones), so a parametric four-engine run lands well under it. Each product costs about 4 prompts on every engine with a key, and products in the same category share their category questions, which are asked once and scored for each product. The MCP tool's default cap is $0.20 per product, which a grounded run will reach, so raise max_cost when you want grounded across four engines. Use --max-cost and start with two or three products.

How long it takes, measured the same way (2026-07-22):

run	measured time
first `npx` (install, once)	about 8 seconds
`audit`	0.3 to 1.7 seconds
`check --quick`, four engines	47 to 51 seconds
`check --quick --grounded`, four	about 97 seconds

The install figure was measured from the packed tarball with an empty npm cache, so a cold npx optifeed-radar audit yourbrand.com from the registry should finish in about ten seconds. A check takes as long as the engines take to answer: it queries several of them across a whole prompt pack, and that wait is provider latency we do not control. check reports live progress while it runs, so you can see which phase it is in rather than watching a blank terminal.

Every run reports what it actually spent, split into setup (brand discovery and prompt generation) and engine calls, so you can reconcile it against your provider bill. Declining at the confirmation prompt still reports the setup cost, because discovery runs before that prompt.

--max-cost 0.20 caps spend. The cap is checked before every call and hitting it returns a partial result flagged as capped, never an error. It is a strong bound rather than an absolute ceiling: an engine's cost is not known until its call returns, so a run can exceed the cap by at most the cost of one unmeasured call per engine. Any overshoot is always reported, never hidden.

Useful check flags: --json (raw envelope), --report report.html (self-contained report), --max-cost 0.50, --quick (smaller prompt pack), --grounded (web-search mode where engines support it), --fail-under 50 (exit non-zero below a threshold, for CI), --yes (skip the cost prompt so an AI agent can run it unattended).

Example

npx optifeed-radar audit example.com      # free readiness score
npx optifeed-radar check example.com --quick --yes
npx optifeed-radar shopping example.com --products "Aria 2, Presto X" --yes
npx optifeed-radar diff example.com        # what changed since last run
npx optifeed-radar sources example.com     # who the AI cited
npx optifeed-radar config                  # which keys are set

config reports only whether each key is present, never the key value.

For shopping, list up to 10 products per run in any order; only the first 10 are checked, and the report is ordered by what the engines did, not by what you typed. A file gives each product a descriptor, which is what rescues an opaque product name - "Aria 2" tells an engine nothing, "quiet home espresso machine" tells it everything:

products:
  - name: Aria 2
    aliases: [Aria II]
    descriptor: quiet home espresso machine
  - name: Presto X
    descriptor: fast dual-boiler espresso machine

npx optifeed-radar shopping example.com --products-file products.yml --yes

FAQ

What is AI visibility? Whether AI engines recommend your brand when a buyer asks them a question, rather than whether you rank in a traditional search index. It is also called generative engine optimization (GEO) or answer engine optimization (AEO).

Is there an MCP server? Yes. The optifeed-mcp server exposes check_visibility, audit_store, generate_buyer_queries, shopping_check and get_snapshot_diff to your AI agents over stdio.

What does it cost? The audit command is free and needs no keys. The check pipeline spends your own engine API credit: measured at about $0.09 for a quick single-engine run, $0.41 to $0.46 across all four, and $0.85 to $1.09 with --grounded. Every run reports what it spent, and --max-cost caps it. You bring your own keys; there is no Optifeed-hosted billing.

Where is my data stored? Optifeed Radar runs locally, saves snapshots on your machine, and has no Optifeed-hosted backend. Each key is sent only to its corresponding AI provider as required to make API calls and is never logged; prompts and responses are handled under that provider's data policies.

Which engines does it support? OpenAI (ChatGPT), Anthropic (Claude), Google (Gemini), and Perplexity. Set any one key to start; set more for broader coverage.

How is the score computed? From sampling real engine answers to unbranded buyer questions, scoring recommendation, position, and share of voice. Scores are estimates and vary between runs. See METHODOLOGY.md for the full method.

For directory maintainers

One-liners for awesome-mcp-servers, ClawHub, and Smithery listings:

Optifeed Radar - Ask real AI engines real buyer questions and score whether a brand gets recommended. CLI plus MCP server, runs locally, BYO keys.
Optifeed Radar (MCP) - check_visibility, audit_store, generate_buyer_queries, shopping_check, and get_snapshot_diff for measuring brand and product AI visibility (GEO / AEO) from your AI agents.

Search intents this serves: AI visibility checker, does AI recommend my brand, generative engine optimization (GEO) tool, answer engine optimization (AEO), ChatGPT brand visibility, AI-SEO, MCP server for brand visibility.

What this does NOT do

It is a point-in-time check, not continuous monitoring.
Scores are estimates from sampling and vary between runs - they are not a guaranteed ranking.
It does not find your products for you. shopping checks the products you name, up to 10 per run; there is no catalog or feed import.
check and shopping spend your own API credit; only audit is free.
It is not a traditional SEO rank tracker.

Catalog discovery (pulling your products from a store or a feed) and feed linting against the Agentic Commerce Protocol (ACP) and the Universal Commerce Protocol (UCP) will arrive in later releases - join the waitlist at optifeed.com.

Status

Under active development, and the repo is public so you can follow along. If this is useful to you, a star genuinely helps. Scores are estimates and say so. Your API keys are used only to call their corresponding providers and are never logged or stored by Optifeed Radar.

License

MIT

More at optifeed.com: https://www.optifeed.com/