The web speaks. Now your agent can listen.

Half the Web, None of the Signal

Agents are text-literate but audio-blind

If you've spent time building an AI agent — a research assistant, a market intelligence tool, a news monitor — you've bumped into the same ceiling. Your agent can read an article. It can parse a PDF. It can search the web and pull back a dozen URLs to scrape and summarize.

But the moment the relevant information exists only as audio, your agent goes quiet. Not because the information isn't out there. The Fed chair gave a speech last Tuesday. The CEO said something interesting on a panel in Davos. A senator made an explicit commitment during a committee hearing. All of it happened. All of it is technically "on the internet." None of it is retrievable by your agent.

Executives say things on earnings calls that never make it into press releases. Scientists describe their findings at conferences weeks before the paper drops. Politicians commit to positions in hearings that don't surface in any searchable text. The gap isn't in what was said. It's in what was indexed.

Why the obvious workarounds fall short

What Sonar is

One tool call. Ranked clips. Citable sources.

Sonar is a search API designed specifically for the AI agent loop. You pass a natural language question — a topic, event, speaker, or claim — and you get back ranked audio clips with accurate transcripts, precise timestamps, and speaker attribution. Your agent can quote them, summarize them, or chain further searches from them.

The index spans news broadcasts, podcasts, radio, earnings calls, social audio, and archives. It updates continuously. You describe what you're looking for; Sonar finds where it was said.

Real scenarios

What this actually unlocks

The clearest way to understand what Sonar enables is to think about specific agents and the questions that would previously stop them.

The API

Designed to drop into your existing agent

Sonar exposes two modes: Retrieve for fast semantic search returning ranked clips, and Research for a full synthesis — a summary with key themes and cited audio evidence. Both are reachable through the same endpoint.

Each result in the response looks like this:

{
  "results": [
    {
      "source":     "Bloomberg Surveillance",
      "source_type": "news",
      "speaker":    "Mary Daly, SF Fed President",
      "timestamp":  "2025-03-18T09:14:32Z",
      "clip_start": 412,  // seconds into recording
      "clip_end":   447,
      "transcript": "We're not seeing the kind of labor market softening
        that would give us confidence to move faster. I think
        two cuts this year is the right framing.",
      "relevance":  0.94,
      "source_url": "https://..."
    }
  ]
}

The response is citation-ready. Your agent gets the source, the speaker, the exact timestamp, and a clean transcript — everything needed to ground a claim in a primary source rather than paraphrasing something that paraphrased something else.

Using it as an LLM tool

Sonar is built to live inside a tool loop. Here's the full integration as a function definition for an OpenAI-compatible tool schema — paste this in and it works:

// Tool definition for your agent
const sonarTool = {
  type: "function",
  function: {
    name: "search_audio",
    description: "Search public audio — news, podcasts, radio,
      earnings calls, social audio, and archives.
      Returns ranked clips with transcripts,
      timestamps, and speaker attribution.",
    parameters: {
      type: "object",
      properties: {
        query: { type: "string" },
        mode:  { type: "string", enum: ["retrieve", "research"] }
      },
      required: ["query"]
    }
  }
};

// Handler
async function search_audio({ query, mode = "retrieve" }) {
  const res = await fetch("https://api.sonarapi.dev/v1/audio.search", {
    method: "POST",
    headers: {
      "Content-Type": "application/json",
      "Authorization": `Bearer ${SONAR_API_KEY}`
    },
    body: JSON.stringify({ query, mode })
  });
  return res.json();
}

What a result looks like in practice

Say your agent is building a brief on Federal Reserve rate expectations. It calls Sonar with "Fed rate cut outlook 2025". Here's what comes back:

Three primary sources, three speakers, three timestamps. Your agent didn't scrape anything, didn't pre-download any files, and didn't rely on a journalist's summary of what was said. It has the actual words, with provenance.

Under the hood

How the index works

Sonar continuously indexes audio from public sources — news networks, podcast directories, radio station feeds, public earnings calls, congressional archives, and social audio platforms. The corpus grows every day.

When your agent asks a question, Sonar searches that index and returns the moments that matter — ranked clips with transcript, timestamp, speaker, and source. Not pages to scrape. Not articles to parse. The actual spoken record, made usable.

The two modes reflect how agents actually use audio search. Retrieve is for when your agent needs raw evidence — the actual clip — to cite or quote. Research is for when your agent needs to synthesize across multiple sources: it finds related clips and returns a structured summary with every claim pinned to its audio source.

What's next

We're building this with early users

Sonar is in public beta. If you're building agents that need to listen to the web — not just read it — we'd love to hear from you at sonarapi.com.

Early users get a direct line to the founding team. We're onboarding a limited number of developers and teams right now, and the people who build on Sonar in the next few months will directly shape what the API becomes. If you have a use case we're not covering yet — a source type, a workflow, a constraint we should know about — we want to hear from you.

Audio is half the web. It's time agents could use it.