Hearsay

A free AI visibility checker, run by the agent you already use.

Watch the full 51-second film with sound.

Give Hearsay your website. Your agent proposes buyer angles, saves a baseline, suggests a page change, and follows the results. Approve its plan or adjust the questions, schedule, and limits yourself. Page versions, answers, decisions, and reports stay in .hearsay/<app-id>/ so your agent can resume later.

Optional Tavily search can use its 1,000 free monthly credits. Six basic searches a day use 180 credits over 30 days, including the baseline, with a proposed 20-credit diagnostic allowance. Your AI assistant's usage follows its own plan. Hearsay reports selected search and AI-agent observations; lead or sales comparisons use the business records you supply. Pricing checked October 2, 2026.

Start with your assistant

Install the Hearsay skill from https://github.com/melonwer/hearsay. Research https://my-app.example and propose a website study to improve qualified leads. Recommend buyer angles, an unchanged comparison page, questions, daily collection, a review date, and usage limits. Offer optional Tavily basic search within 200 credits for 30 collection days. Save page version 1 and show me the actual plan before enabling collection. Keep an evidence-linked history, analyze each collection, and propose one focused change for my approval.

The agent chooses defaults and you can edit them. The first review can happen after one day; judging leads or sales depends on exposure, traffic, and supplied outcomes. Changing a page does not make new text immediately available to search-backed agents. See the study workflow.

For manual installation, copy the complete skill/ directory into your host's skill directory, for example .agents/skills/hearsay for Codex or .claude/skills/hearsay for Claude Code. Keep the references, templates, and schemas. Initial host research needs no Hearsay server, Node installation, API credits, or extra account.

Hearsay researches the product and discovers direct and adjacent competitors with supporting links. It proposes evidence-linked improvements and separate drafts. Applying a patch or publishing content requires your request.

Independent trials start with six neutral buyer questions and one answer per selected route. Fresh measurement sessions receive only neutral questions. Without fresh sessions, the report is a research audit with no independent visibility rate. Repeats preserve questions and revisions, show exact sample counts, and explain comparison eligibility.

See installation, upgrades and removal. Build the portable archive, Codex plugin, Claude Code plugin and Gemini CLI extension from one canonical skill:

node scripts/build-bundles.js
node scripts/verify-skill-discovery.js

The second command checks installation and discovery in isolated host configurations without inference. Archives live in dist/ and contain no local database, credentials or reports.

Optional runtime and dashboard

With Node 22.13 or newer, node bin/hearsay.js --help exposes agent discovery, bounded panel execution, website study history, optional Tavily collection, outcome imports, saved reports, and schedules. npm link installs the local hearsay executable. These commands need no HTTP server. See the CLI guide and study commands. Runtime collection marks assistant analysis due. Automatic analysis requires a connected host agent task; an approved configuration alone does not install one. Tavily strict free mode also requires a provider limit that prevents paid overage.

Codex and Claude Code account routes use existing logins and keep results separate. Extra routes are selected explicitly; API keys are removed from account invocation environments. Costs and remaining allowance stay unknown unless reported. Antigravity agy-cli replaces the planned Gemini CLI measurement route. Its adapter extracts observed web searches from the event stream and preserves other tool activity in the trace. Tool inventories may differ between installations. The Gemini extension still installs the portable skill. Gemini API grounding remains a separate option.

Import saved evidence through hearsay import <evidence.json> --database /absolute/path/hearsay.db, the dashboard's Research page, POST /api/research/import, or hearsay_research_import. List/detail operations are GET /api/research, GET /api/research/:id, hearsay_research_list and hearsay_research_get. Imports retain external provenance and never enter native measurement series. Research recommendations can be proposed in Opportunities, where acceptance requires a separate review.

The dashboard and API workflows are optional. Their setup follows below.

What it does

Buyers ask AI tools which product to buy. Hearsay records how the specific API or signed-in agent route you select answers your own reviewed questions.

Hearsay asks your buying questions several times and counts how often you get mentioned or receive a positive recommendation. Answer rates show a confidence interval and the number of comparable answers behind them. Retained answers let you inspect the text and evidence behind each selected series. An explicitly enabled Codex agent or Claude Code agent runs through its authenticated local CLI. Those measurements retain final answers, provider-confirmed web-search events, citations and redacted event artifacts. They are sibling evidence to API runs, never a blended score, and they consume the signed-in plan allowance or possible overage — subscription runs are not free or unlimited.

Direct API keys are optional. Add keys for OpenAI, Anthropic, Gemini and/or Perplexity when you want direct API measurements and expanded coverage. API providers bill those questions separately, usually by usage. Hearsay's built-in quote and computed usage cost cover API calls only; they do not estimate or report subscription allowance consumption.

OpenAI API runs can use live web search on gpt-5.6-luna. Set HEARSAY_OPENAI_SEARCH_POLICY=auto or required to use the Responses API route; the default off keeps the existing Chat Completions request. Search runs always show a quote before starting. It forecasts one web-search call per target, but the hosted tool has no enforceable internal call ceiling. The daily API schedule stays search off until you separately confirm POST /api/search-schedule with a target_ceiling and its returned quoteId. Changing the execution profile or exceeding the target ceiling suspends scheduled search. Use DELETE /api/search-schedule to remove recurring search consent. No live provider smoke test is run automatically.

Anthropic API runs can use the basic server-side web search tool on claude-sonnet-5 with HEARSAY_ANTHROPIC_SEARCH_POLICY=auto. Hearsay limits search to three calls per target and one continuation within the target timeout. The auto policy includes complete answers where Claude chooses not to search. The same on-demand quote and separate daily-schedule consent apply. An auto answer without search is comparable only when the provider response confirms that no search ran. A required answer needs a completed search event. A missing query string is shown as unavailable, never reconstructed from an answer or source URL. Gemini API search is off by default. Set HEARSAY_GEMINI_SEARCH_POLICY=auto on gemini-3.6-flash to enable Google Search grounding through GenerateContent. The grounded route has its own execution profile, needs the same run quote and schedule consent, and has no enforceable internal search-call ceiling. Hearsay retains the original grounding metadata and shows Google's Search Suggestions in an isolated frame beside a grounded answer. If Gemini omits the Suggestions after a search, Hearsay records a failed target without displaying its grounded content. Search billing remains partial when the API does not report the number of billable calls. Perplexity Sonar's historical built-in search remains a legacy profile without verified query wording.

Set up the optional dashboard by asking

If you use a supported paid Claude Code account, paste this into Claude Code and it can do the whole setup:

Install and set up https://github.com/melonwer/hearsay for my brand, [your brand], and track it against [competitor] and [competitor]. Skip the demo data. Show me the questions before you save them.

It clones the repo, starts Hearsay, connects itself to it, drafts the buying questions worth tracking, shows you that list before saving anything, tells you what a run will cost before spending a cent, and then reports back.

This works because Hearsay speaks MCP (Model Context Protocol), the standard way an AI assistant plugs into an outside tool. MCP lets Codex and Claude Code operate Hearsay; it does not make an arbitrary MCP client an inference backend. Hearsay inference runs come only from the explicitly enabled Codex and Claude Code CLIs or from optional direct API keys.

<details> <summary>The exact steps, if your assistant gets stuck</summary>

git clone https://github.com/melonwer/hearsay && cd hearsay
node --version                        # must be 22.13 or newer

cp .env.example .env                  # subscription flags and optional API keys go here
node server.js > hearsay.log 2>&1 &   # redirect, or a backgrounded server hangs the shell
curl -s "http://127.0.0.1:$(tr -d '\\n' < data/hearsay.port)/api/status"   # repeat until this returns JSON

# Add the MCP server for the client you use (see the subscription sections below).
codex mcp add hearsay -- node "$PWD/mcp/server.mjs"
claude mcp add -s user hearsay -- node "$PWD/mcp/server.mjs"

Hearsay's tools appear only after the assistant reconnects. Quit Claude Code and start it again, then say "carry on setting up Hearsay". If they still do not show up, work over the JSON API instead, which needs no setup: GET /api/status, POST /api/setup/drafts, GET /api/setup/drafts/:id/review, POST /api/setup/drafts/:id/approve, GET /api/cost/estimate, POST /api/run, GET /api/runs/latest, GET /api/series, GET /api/series/summary, GET /api/series/evidence?series_id=...&intent_id=..., GET /api/answers?series_id=..., GET /api/alerts. GET /api/summary?days=30 remains the legacy API view. There is no index page at /api.

The Evidence page groups the buyer questions, observed search queries, source observations, final-answer citations, and answer receipts for one selected measurement series. Open a group's supporting receipts to inspect the recorded action and source links. Query counts describe this configured panel, not customer search demand. Manual theme labels are separate from the provider's wording. The evidence report guide explains the denominators and unknown associations.

The Opportunities page turns selected receipts and repeated patterns into a reviewable shortlist. Each observed finding links to the answer and evidence IDs in one exact series and window. You can prioritize, dismiss, combine, o