What's new

Agents can now be given a knowledge base of documents, web pages, and text to draw on in conversation.

• Upload PDFs, Word documents, images, or text files, add a web page, or import a whole site from its sitemap. Everything is parsed to text and organized in folders on the Knowledge Base page.
• Attach documents to an agent from the Knowledge tab of the builder. A meter shows how much of the prompt they use.
• Manage documents through the API, and swap an agent's knowledge base per session with overrides.

Check it out in the platform dashboard or read the docs.

Agent sessions now have optional transcripts and audio recording, through both the dashboard and API.

• Open a session under Usage & Sessions to read the transcript, with the agent's response time on each turn, and play back or download the recording.
• Fetch sessions, transcripts and audio from the API, or have the transcript sent to a webhook when a call ends.
• Recording is off by default and set per agent. You choose how long records are kept, from not at all to a year.

Check it out in the platform dashboard or read the docs.

Agents are now first-class in the Keyframe API.

• Agent IDs, short-lived session tokens, and public or private agents. Publishable keys still work.
• Dynamic variables in system prompts and greetings.
• Empty greetings, so the agent waits for the user to speak first, and interruptible greetings.

Check it out in the platform dashboard or read the docs.

See how your agents and avatars are being used. Sessions, minutes and peak concurrency over any window from the last 24 hours to the last 12 months, broken down by avatar or API key, with a searchable session history.

Find it under Usage in the platform dashboard.

Teams can now share a workspace in Keyframe. Invite your team to an organization, set each member's role, and collaborate on agents, avatars, API keys and more.

Create one from the platform dashboard.

persona-2.0-live, the world's most expressive, controllable avatar model, is now in preview.

• Word-level control over avatar emotion in live conversation
• Industry leading realism
• 1000s of concurrent connections; $0.06 per minute

Try it on our home page — talk to it and watch each tag light up as the face changes.

persona-1.5-live has graduated from early access and is now generally available to everyone. Create an avatar with it and embed it in a website in minutes, or drive it through our API with your own agent stack.

persona-1.5-live, a 200M parameter model that sets a new standard for realism and efficiency, is now in early access.

• ~180 ms from audio in to video out; typical turns land around 500 ms end to end
• Emotion as an explicit conditioning signal, steerable mid-conversation
• $0.06 per minute, with one GPU serving many concurrent streams

Read the research post.

persona-1-live, a real-time streaming variant of persona-1, is now in early access — capable of holding dynamic, human-feeling conversations with low latency.

• Next-token prediction scheme optimized for interactive video generation
• No loss in persona-1's speed, expressive fidelity, or cost
• Pricing comparable to persona-1: just a few cents per minute

Read the research post.

persona-1, our expressive, faster-than-realtime talking head model, is now in early access.

• ~2x real time on last-gen consumer hardware (RTX 4090), ~4x on L40S
• Up to eight live streams in parallel on a single low-cost enterprise GPU
• A few cents per minute — ~1/2 the cost of the cheapest competing models

Read the research post.