app.llmaudit由 llmaudit.app 发布官方发布者v1.1.0更新于 Nov 2, 2026
Measure whether AI assistants actually recommend a brand, from measured answers.
测量 AI 助手在买家寻求推荐时是否真的推荐某个品牌，并报告实测回答而非模型自述。
Web executable，通过 Streamable HTTP。 远程服务在工作区中配置后即可从网页运行时运行。
把这段提示词粘贴到 Claude Code、Codex、Cursor 等智能体，或 SourceWeft 对话中。智能体会读取该服务器的安装指南，先向你展示发布者、端点和登录方式，经你同意后再把它添加到你的客户端。密钥不会写进任何命令。
阅读 https://sourceweft.com/mcp/app.llmaudit%2Fllmaudit/install.md，并按其中说明把这个 MCP 服务器接入我的客户端。修改任何配置之前，先告诉我它的发布者、端点和登录方式，等我确认。不要把 API 密钥写进命令或我会提交的文件；问我密钥应从哪里获取。把它添加到你客户端的 mcpServers 配置中。
{
"mcpServers": {
"llm-audit": {
"type": "http",
"url": "https://mcp.llmaudit.app/mcp"
}
}
}MCP server for askedthrice.com. Lets anyone's assistant measure whether a brand shows up when buyers ask an AI for a recommendation.
Closed since 2026-10-29. Asked Thrice no longer runs free measurements; the product is a paid scan that needs an account.
start_visibility_checkreturnsstatus: "closed"with the link to askedthrice.com and does not call the API.get_visibility_checkstill reads measurements started before that date. The sections below describe how it worked while it was open.
Hosted at https://mcp.askedthrice.com/mcp (streamable HTTP). Listed in the official MCP registry as com.askedthrice/askedthrice. The stdio bridge for clients that only launch local commands is the npm package askedthrice-mcp, in npm/.
Full spec: specs/features/geo-13-mcp-server.md in the private llm-rank-tracker repo.
or, for command-only clients:
start_visibility_check starts the measurement and returns a runId.get_visibility_check collects it. Returns running until the providers answer.askedthrice://methodology explains how it is measured.Two tools rather than one because a measurement takes up to two minutes: the audit is 15 to 60 s and the measured map another 15 to 60 s.
It returns the measured map: the buyer questions that were asked live and who won each answer. Never the providers' self-report, which is what each model believes about the brand. The two numbers drift apart a lot in practice, and the self-report is always the optimistic one; that is why the tool contract does not expose it in any field.
It never returns a 0 to 100 score: a verdict in three bands and counts.
Every free measurement spends askedthrice's API keys, on the order of USD 0.05 to 0.15. Three guards, all covered by tests:
MCP_MONTHLY_RUN_LIMIT (default 200) as the channel ceiling, separate from the web funnel budget.npm run stats reads mcp_runs and reports how much of the monthly ceiling is left, who called, which domains were measured and which one currently holds the 30-day quota.
Needs SUPABASE_URL (or NEXT_PUBLIC_SUPABASE_URL) and SUPABASE_SERVICE_ROLE_KEY. --days is the daily detail, --history how far back it looks; the current month is always included in full, because that number has to match the guard's.
The first thing it prints, when there are any, is the orphaned runs: reserved rows that never got their audit_id. They return "Unknown runId" to the client forever while holding their domain's free quota and counting against the monthly ceiling. Delete them by run_id.
Vercel Web Analytics cannot measure this: it works by injecting a script into HTML, and here there is no HTML and no browser to run it, every request is a JSON-RPC call from an MCP client. It would read zero forever.
Dockerfile builds the HTTP server and serves it on 0.0.0.0:3001, which is what Glama and similar directories run to verify the server starts and answers introspection. Port and host are read by xmcp.config.ts at build time, so they go as ENV before npm run build, not at docker run. Without SUPABASE_* the run store is in memory: fine for a check, not for serving traffic.
Tests run against real fixtures in fixtures/ (already-paid runs from prospecting). None of them touch the network or call a provider.
server.json is published with mcp-publisher using DNS auth on askedthrice.com. The key lives outside the repo as a PEM; the CLI wants the raw ed25519 seed in hex:
Bump version in server.json on every publish; the registry rejects a repeat. When the npm package changes version, update packages[0].version too.
来源：README.md，提交 0c5e67c
Get visibility check
Collect the result of a measurement started with start_visibility_check. Returns status running while the providers are still answering, so poll every 20 to 30 seconds. When ready it reports a verdict, how many of the buyer questions the brand won, which competitors were named instead, and which providers actually answered.
runId必填Start visibility check
Measure whether a brand shows up when buyers ask AI assistants for a recommendation. Asks real buyer questions to OpenAI, Anthropic and Gemini and reports what they answered, not what they claim they would answer. Free, no signup, one measurement per domain every 30 days. Returns a runId: call get_visibility_check with it after about a minute. If the domain was already measured this month the response says quota_reached and includes when it ran.
brand必填website必填category必填languagelocationcompetitors已满足适用检查项的 100%
该评分衡量此收录条目的完整程度，依据注册中心条目、端点与公开来源计算得出。它不是对服务器本身的评价，也不表示服务器安全或当前可用：可连通性是带日期的检查结果，评分也不包含安装量或用户评价。
{"mcpServers":{"askedthrice":{"url":"https://mcp.askedthrice.com/mcp"}}}{"mcpServers":{"askedthrice":{"command":"npx","args":["-y","askedthrice-mcp"]}}}node --env-file <file.env> scripts/stats.mjsnode --env-file <file.env> scripts/stats.mjs --days 30 --history 180npm run dev # xmcp devnpm test # vitest, no networknpm run build # xmcp buildHEX=$(openssl pkey -in <key.pem> -noout -text | sed -n '/^priv:/,/^pub:/p' | grep -v '^priv:\|^pub:' | tr -d ' :\n')mcp-publisher login dns --domain askedthrice.com --private-key "$HEX"mcp-publisher publish server.json