Fish Agent

• Answering a support line that looks up orders and issues refunds through webhook tools, then warm-transfers hard cases to a person with a spoken briefing
• Taking appointment bookings and restaurant reservations by phone against your own scheduling API
• Putting a talking agent on a website with the drop-in widget or inside a web app with the React SDK
• Placing reminder or feedback calls from your backend with per-call dynamic variables
• Adding an AI agent to existing numbers from Twilio, Asterisk or another SIP carrier, or to a WhatsApp Business number
• Running an agent on your own fine-tuned or proxied model through a custom OpenAI-compatible endpoint while Fish Audio handles speech and telephony

Fish Agent is a voice company's bet that turn-taking belongs inside the speech model. In English calls, Live 3's ability to say 'mm-hmm' through a read-out phone number, keep talking through a 'yeah' and stop for a real 'wait, actually' targets the moments where cascaded agents most often feel robotic, although the 186-millisecond reaction time and the quality case are Fish Audio's own claims with no independent benchmark yet. Around the model, the platform already covers what a production phone agent needs: versioned agents, webhook and client tools, SIP and WhatsApp numbers, warm transfers, LLM-judged tests and post-call analysis, at an all-inclusive $0.07 a minute (half that until the end of 2026) that is easier to budget than stacks billing speech, LLM and platform separately. The gaps show its stage: it is still a beta, Live 3 is only days old, knowledge bases take only text files, outbound is one API call per dial, and non-English agents lose the full-duplex advantage. It fits English-first teams that want natural phone conversation at low cost and can live with beta churn; teams that need provider freedom should compare Vapi, those that need many languages, native mobile SDKs or a longer track record should look at ElevenLabs Agents, and those running outbound campaigns should look at Cartesia Managed Agents' batch calling.

About Fish Agent

Fish Audio, the text-to-speech and voice-cloning company operated by Hanabi AI Inc., has run its agent platform as a public beta open to every Fish Audio account since September 2026, and in November 2026 added Live 3, its full-duplex speech model, which new agents now start on. You build an agent in the Builder console or over the REST API: a system prompt and first message, a voice from Fish Audio's library or your own clone, a speaking language, knowledge sources, and tools. Webhook tools call your HTTP endpoints during a conversation (you can import them from a cURL command and mock responses for testing), client tools run functions in your web app through the SDK, and system tools let the agent hang up; phone agents can also transfer to people, cold or warm with a private spoken briefing. Edits land in an autosaved draft, and publishing freezes it into a numbered version, so live calls never run half-finished changes. The same agent can then sit behind a two-line website widget, the @fishaudio/agent-client and @fishaudio/agent-react SDKs, a purchased US or Canadian number, a number imported from your own carrier over SIP, or a WhatsApp Business number; outbound calls are placed one at a time through the API. Every session keeps a transcript, a tool-call timeline and (by default) recordings, and an LLM pass after each call writes a summary, extracts fields you define and scores success criteria, delivered by signed webhooks. Before publishing you can run single-turn, tool and full simulated-conversation tests judged by an LLM. What sets it apart is the speech layer: new agents start on Live 3, which listens while it speaks and decides itself when to say 'mm-hmm', keep going or stop, instead of waiting for silence like the usual speech-to-text, LLM and text-to-speech cascade. Live 3 speaks English only, though; agents in any of the other 51 supported languages fall back to the cascade on Fish Audio's Drama 3 or S2.1 Pro voices. Pricing is simple by voice-agent standards: $0.07 a minute covers speech recognition, voice, knowledge retrieval, tool calls and a set of basic LLMs, cut to $0.035 until December 31, 2026, with phone calls adding $0.01 a minute and numbers $1.20 a month. The tradeoffs are those of a young beta: features and APIs may change, knowledge sources are limited to Markdown and plain-text files, managed numbers cover only the US and Canada, there is no batch dialer, and compliance options such as a HIPAA BAA, zero data retention and SSO are reserved for the sales-led Enterprise plan.