Sourceblender builds infrastructure for AI agents that work together over time: shared memory, an agent runtime, and real-time voice. The memory layer is open source today; the runtime and voice are running in-house, with public releases planned. It is built and run every day on our own hardware.

Built by Eric Mey.

A memory server for a small fleet of agents, so that what one learns the others can use. It has three memory planes, runs local inference, and has a lifecycle engine that matures raw captures into a knowledge base a person can review. Host adapters connect agents to Musubi. For the hosts it supports, the shared harness supplies turn capture and verified delivery. Retrieval quality is measured by an eval harness that runs nightly in CI.

An OpenAI-compatible endpoint that routes each turn to a specialist agent (conversation, image, audio, search, video), each with its own prompt and tools. Clients see one model; they never see the internal tool calls. In internal daily use. Private · public release planned.

A private LiveKit plugin in development: spoken conversation with agents, in real time and over the phone, built with full-duplex turn-taking.. Private · public release planned.

• ComfyUI Immich: ComfyUI nodes that upload images to Immich, with metadata and album management.
• ComfyUI Kotodama: a ComfyUI node that enhances text-to-image prompts through any OpenAI-compatible endpoint.

Two separate things: who can see it, and how far along it is.

• Visibility: public (source open) or private. A private project may add public release planned when we intend to open it.
• Lifecycle: seed (a placeholder repository), proof of concept, developing, released (published through GitHub Releases or a package registry, named where it matters), maintained, or archived.

Start with a project's README for installation and support. Report security issues through that repository's SECURITY.md, or the organization default.