Models, agents, tools, and the mesh between your machines, running on devices you control. Cloud only when necessary. No per token bill.
Up and running in under 5 minutes.
Files, shell, APIs, skills, subagents: an agent that carries long multi-step work end to end, with you approving what executes.
How agents work
Catalog → quantization → pull → loaded on your GPU. Aliases let apps ask for qwen while the router picks the model.
Attach the inference servers you already run and register keyed cloud providers. Their models land in one list, behind one router, with real measured tok/s.
How providers work
One Try button from any model to a raw conversation. Thinking streams in, tokens-per-second on every turn, nothing saved. Need the ceiling instead of a single turn? The built-in benchmark swarms the box to measure how many agents it can host.
About the Playground
A swarm of real agents with real sessions, driven against your instance until it bends. Density, completion rate and the breakpoint, measured by running it on your hardware, not read off a spec sheet.
About Benchmark
mimOE nodes discover each other. The workstation's 35B model shows up on your laptop, usable by name.
How the mesh works
Agents don't wire their own MCP servers. Attach any MCP server once, scope it with client keys, and every tool call flows through one governed MCP gateway.
How the MCP gateway works
Switch to another instance in one click. If its IP changes, Studio finds the node on the mesh and follows it.
How remote connections work
Each call through the router becomes an OpenTelemetry trace, capturing model, routing, tokens, timing, and content on-device. Governance, the policy and audit layer, is being built on exactly these records.
How trace & governance work
Models, agents, tools, mesh, and the receipts to prove what happened.
One-click runtime install includes the engine, model store, router, and a tool gateway. Works offline.