PodAgent is a voice-first podcast playground built for Hack #1: Firecrawl x ElevenAgents.
The idea is simple: you are the host, an ElevenLabs voice agent is your guest, and Firecrawl gives that guest live web context during the conversation. Instead of a chatbot or search box, the product feels like recording a short podcast with an AI co-host that can look things up in real time and then turn the session into something playable and exportable.
podagent.MOV
PodAgent is a 3-step flow:
- Start with a topic
- Talk to an ElevenLabs voice agent live
- End the session and review/export the conversation
During the live session:
- the user speaks first
- the ElevenLabs agent responds in voice
- Firecrawl powers live search through a webhook tool
- the transcript updates in real time
On the export screen:
- the full conversation audio is fetched back from ElevenLabs
- the session can be played back immediately
- a vertical cover video is shown as the visual layer
- the final MP4 is rendered by combining the cover video with the conversation audio
Most voice agents are built like assistants. We wanted to explore a different interface: a podcast format.
That makes the interaction feel more natural for:
- interviews
- debates
- commentary
- explainers
- creator workflows
The key product idea is that the AI guest should not just sound good. It should also be able to research before it speaks.
That is where the Firecrawl + ElevenLabs combination becomes interesting.
The app uses the ElevenLabs React SDK to start a live voice session with an existing ElevenLabs agent.
We use:
useConversationfor the real-time conversation loop- client tools for lightweight actions like clip bookmarks and sound effects
- contextual updates so the topic can shape the session without relying on unsupported prompt overrides
The ElevenLabs agent is configured with a search_web tool that points to our backend webhook.
That webhook:
- receives the search query from the agent
- calls Firecrawl Search
- returns structured live web results back to the agent
This gives the conversation a clear rhythm:
- host asks
- agent searches
- agent answers
When the conversation ends, the app stores the ElevenLabs conversation ID.
From there:
- the export screen fetches the full conversation audio from ElevenLabs
- playback works directly inside the app
- an MP4 export route uses FFmpeg to merge:
- the uploaded thumbnail/cover video
- the ElevenLabs conversation audio
- the title overlay rendered into the video
- Next.js 15
- React 19
- TypeScript
- ElevenLabs React SDK / ElevenAgents
- Firecrawl Search API
- FFmpeg via
ffmpeg-static
The UI is intentionally minimal and tool-like.
We pushed it toward a clean white workspace aesthetic:
- flat surfaces
- low chrome
- grayscale UI
- red only for live/recording actions
The goal was to make the conversation and export flow feel like a working product, not a marketing landing page.
- topic-based session start
- live ElevenLabs voice conversation
- Firecrawl search integration through an agent tool webhook
- real-time transcript updates
- export screen with interactive playback
- custom vertical cover video support
- MP4 generation from cover video + conversation audio
PodAgent treats voice AI as a media format, not just an assistant.
Instead of asking a question and getting an answer, the user gets a conversation they can listen back to, package, and eventually share.
That is the core experiment behind the project:
What happens when live web research and conversational voice agents become a podcast workflow?