Get to your first banter in under a minute.
A short, illustrated walk-through of every screen — followed by the field-by-field reference and a troubleshooting list.
Quick start
-
1
Just open the site. There is nothing to install, no account to create, and no API key to paste — voice generation runs on a server-side key, so you can start listening straight away.
-
2
Open the composer. On the home feed, click the big "Create a banter" card. The form expands with two agent slots, a topic field, and optional category + tags.
-
3
Hit Listen now. Either fill the fields by hand or click "Draft with AI from one line" and let the model populate everything from a single prompt like "Gandalf and Dumbledore on student safety." Then press the play button. Audio starts in a few seconds.
Accounts & privacy
Agent Banter uses Google's Gemini Live API for voice generation, billed to the site — you never need your own key, and there is no signup wall.
- Everything works signed out. Your banters, transcripts, agents, likes, and comments live in your browser's localStorage.
- Signing in with Google is optional. It stores only your name, email, and profile picture so the site can greet you.
- Because your library is browser-local, it does not follow you between devices — and clearing site data clears it.
- Voice sessions connect with short-lived, single-use tokens issued per call, so no long-lived credential is ever exposed to the browser.
Composer fields
Every field is optional except agent descriptions + topic. Here's what each one does:
Agent A / Agent B — Name
Short label shown on the live badge + transcript. If you leave it blank, we'll infer it from the description ("named Rook…" → Rook).
Agent A / Agent B — Description
The persona. Tone, beliefs, mannerisms, what they care about. 2–4 sentences is plenty. Strong personas with sharp disagreement make the best banter.
Voice (per agent)
Each agent needs a different voice — that's how you tell them apart by ear. The composer warns and disables Listen now if both are the same.
Topic
A single sharp question or framing. The narrower it is, the better the banter. "Is Postgres still the default in 2026?" beats "tech stuff."
Emoji + Title
Shown on the feed card. Both optional — if you skip the title, we use the first 80 chars of the topic.
Category + Tags
Category is free text — pick what fits ("Tech", "Philosophy", "Comedy"). Tags are chips: type, press Enter or , to add, × to remove, capped at 6.
AI quick-start
Don't want to type six fields? Click the small ✨ Draft with AI from one line chip at the top of the composer. A single sentence drafts the entire banter.
Good seed prompts:
- Two named characters — "Gandalf and Dumbledore on student safety"
- A clash of stances — "Stoic philosopher vs. Silicon Valley founder on work-life balance"
- A topic-only seed — "Whether AI art counts as art" (we'll invent two opinionated personas)
Everything the AI drafts is editable — change names, swap voices, tweak personas, rewrite the topic. The "✓ Drafted from …" note shows what seed was used.
Listen, save, publish
Three actions, three meanings:
▶ Listen now
Starts the live audio call. Both agents connect, take turns speaking, transcript writes itself.
Publish
Saves the talk to the feed (locally, with a public badge) without listening — anyone who opens the card can start it themselves.
Clear
Resets every field to defaults. Disabled if the form is already empty.
While the agents are talking — or after they stop — you'll see Save and Save & publish in the transcript footer. Both snapshot the transcript so reopening the talk shows the same conversation.
Other modes
Banter (two AI agents) is the main attraction, but Agent Banter ships a few sibling modes:
- Chat — talk one-on-one with a single agent. You type, they reply with voice.
- Solo — one agent monologuing on a prompt. No second speaker, no input from you.
- Voice — a real-time voice call with one agent (you speak, they reply).
- Modes — a directory of all current + planned conversation formats (panel, debate, interview, roast, etc.).
All modes share the same agent library and voice list.
Agents library
Personas you want to reuse can be saved to the Agents page (top-right in the navbar). Click New agent, give it a name, voice, and persona, save. The library lives in your browser's localStorage — there's a small read-only seed pack shipped by default.
Saved agents don't auto-populate the composer yet — for now, copy the name + persona into the composer fields by hand. That wire-up is on the roadmap.
Troubleshooting
"Listen now" is greyed out
Three causes: missing agent description, missing topic, or both agents using the same voice. The form labels turn rose-tinted on the field that's the problem.
"Voice service unavailable" or auth errors
Voice sessions use short-lived tokens issued by the site. Reload the page to get a fresh one. If it persists, the upstream model may be briefly rate-limited — wait a minute and retry.
No audio after "Connecting…"
Check your browser allows audio playback (some browsers block autoplay until you interact with the page). Click anywhere on the page first, then hit Listen now.
My talks disappeared
All talks, transcripts, likes, and comments live in your browser's localStorage. Clearing browser data clears them. There is no cloud backup — signing in does not sync your library.
AI quick-start fails to generate
Usually a rate limit or transient model error. Wait a few seconds and try again — the site enforces a fair-use limit per visitor.
Still stuck
Open Contact, describe what you tried + what happened + what browser you're on. Real humans read every message.