Livepeer

AI APIs price work in three common units: per token, per second or hour, and per request. Here's how AI companies track it.

An AI agent has five parts: a model, instructions, tools, memory, and a loop that plans, acts and checks its work. Increasingly, tools connect through MCP. For images, video and voice, the open-source Livepeer Agent lets your agent plan and make media through MCP, today.

Anime to live action, made on Livepeer. @harendrashakya_ shows how MangaFable turns a story into manga, animation and live-action shots, built with the Livepeer Agent MCP and backed by an early SPE micro-grant from @GeoDeFine.

The best defence against AI vendor lock-in: build so your app can always move. With Livepeer's Live Runner, a node on an open network hosts your app with no rewrite. One sample app serves an open model in a common AI chat request format, so existing chat code works unchanged.

Treasury diversified, Embody bugs fixed, and a first look at the M3 agent workspace. @GeoDeFine's SPE update from Watercooler: agent telemetry via OpenTelemetry and Arize Phoenix, evals and datasets over MCP, and many agent runtimes in one interface.

Open-source LLM doesn't mean free LLM. Here's why:

The Livepeer Explorer, rebuilt from scratch. @AdamSoffer walks through the new version at this week's Watercooler: a portfolio view with historical rewards and fees, multi-wallet tracking, stake transfers in the app, and a live treasury view.

With Live Runner, a Livepeer orchestrator hosts your app as it is, from a web service to live video. The Livepeer part is a handful of clearly marked lines.

From a tap to a GPU, a Livepeer request makes four open stops.

Self-sovereign means no single company holds the off switch. Livepeer's open builder stack is designed so you install, run, upgrade and restore it yourself, with your own data and credentials.