← OpenAI
Creator Magic
GPT-Realtime-2 Voice Agent: I Build It Live
⏱ 56 min video · 3 min read18 Jun 2026
TL;DR
The creator live-builds a voice AI assistant called Jarvis using GPT Realtime 2 (OpenAI's model released ~6 weeks prior) and the Tank framework, connecting it to Zapier MCP for tool access and Imagen via Nano Banana 2 for real-time image generation. The session is raw and experimental, revealing both the promise and current limitations of voice-driven AI agents.
Key points
1
GPT Realtime 2 was released by OpenAI roughly 6 weeks before this stream and features a 128,000 token context window, GPT-5 class reasoning, and multimodal support (text, image, audio) with function calling.
2
The Tank framework (creator's own tool) orchestrated Claude Code to scaffold a LAN-hosted Jarvis voice assistant from a single prompt in under 6 minutes, including a Caddy reverse proxy for microphone HTTPS access.
3
Zapier MCP integration partially worked — the voice agent could detect available Notion actions — but failed to actually create pages, highlighting real-world friction with MCP-connected voice agents.
4
Swapping to Nano Banana 2 (Google Imagen) for image generation via voice worked impressively: the user described scenes verbally and Jarvis generated images in real time, useful for thumbnail ideation.
5
Tank now supports Fusion mode — synthesizing answers from multiple models simultaneously (e.g. local Qwen 3, GPT-OSS 120B, and GPT 5.5 via API) — and has added scheduled tasks, keyboard shortcuts, Linux/Proxmox support, and community pull requests.
Actionable insights
→
To use browser microphone access on a LAN-hosted app, route it through HTTPS via a reverse proxy like Caddy — plain HTTP blocks mic permissions in modern browsers.
→
When building a voice agent with GPT Realtime 2, add function-calling definitions rather than relying on MCP connections in the playground, as MCP wiring is unreliable in that environment.
→
For rapid visual brainstorming (e.g. YouTube thumbnails), a voice-to-image pipeline using GPT Realtime 2 + an image model is already usable today — describe a scene verbally and iterate hands-free.
Notable quotes
“I kept telling me it can't make an image of Mario, but it made an image of Mario. So there's definitely a wrinkle there somewhere.”
“Just in the last day or so, Claude Opus 4.8 has got a lot dumber. Like, it's taking a lot more time, it's making errors and it's having to go around again a couple of times before it gets things correct.”
“It shows you how easy it is even to this very day with GPT 5.5 to jailbreak even inadvertently and get a result.”
Worth watching?
⏭️
Worth watching the full video?
Skip the full video unless you enjoy unedited live-coding streams — the key findings, limitations, and build steps are all captured here, and the video contains significant dead time waiting for Claude to code.
Saved you some time? The creator still deserves a like.
Want this for your own channels?
Add the channels you follow. Every new video summarised and in your inbox the moment it drops. From £4/month.
Try it free