Agentix-Native Workstation
Built on mimOE, mimOE Studio is an Agentix-Native Workstation that shows running agents, their traces, and services, and exposes mimOE capabilities so you can interact with them directly.
Build, test and measure your agent on one engine
Builder Bench, Agent Harness, and Workflow & Compute Benchmarking. Together as mimOE Studio, powered by the mimOE Operating Engine. The agent you build is the agent you test and measure, on hardware you control.
Pick any model this instance can reach, name the agent, give it a sandbox directory. It deploys as its own mim with its own API path.Work with the agent on real tools. Risky calls wait for your approval, and every action is traced on the device that served it.How many agents can one box host? Launch N serverless agents against the one you built, on the real model, and watch each one verify itself.Extend the three panes, or bring your own on the same engine.
An animated mimOE Studio window. First, the New agent wizard picks an on-device model, names the agent Field inspector, gives it a sandbox directory and deploys it. Then the agent is asked to summarise today's reports: its bash call goes through the MCP Gateway and waits for approval, is allowed once, it replies, and the exchange is traced on the device. Then Benchmarking launches forty serverless agents against it on the real model, four holding the GPU at a time, and all forty verify. Last, your own app calls the agent's API path and the qwen alias over the OpenAI-compatible API, and each call appears in Studio's Trace as it lands. Under the window, all four panes sit on one mimOE engine of 10 to 20 MB, agnostic to language, OS, device, network, cloud and LLM.
Every node, every model, one surface
The Workstation is the visual control and management surface built on mimOE.
- Load, unload and switch models on demand, between My Device and Network
- Benchmark CPU against GPU and pick the most efficient execution path
- Validate any model or agent live, via chat or direct API, before anything reaches production
- Mesh-wide visibility and live workload management
- Trace every call end to end: routing decisions, latency, inputs, outputs, token usage
- Watch live runtime metrics, memory included, from desktop or mobile in the field
Everything your AI needs to work, in one place
The Agent Harness is what turns a smart model into a capable one. Your model brings the thinking. The harness gives it a place to act, a memory that keeps up, a way to check its own results and rules it always follows.
In mimOE Studio, it all runs on your own computer, with the model you choose, and you have the final say.
Measure it by running it
The Benchmark is a stress test for your AI setup. It fills your machine with real agents working at the same time and measures exactly how much it can take. Every result is checked, so you can trust the numbers. And it runs separately from your real work, so your actual agents and their history stay untouched.
Real Density
How many agents your box sustains for a given workload, measured by running them.
Verified Completion
Output checked against the expected result of the task, for every executed agent.
The Breakpoint
Sweep the fleet size upward and watch where the run bends: queue depth, agents in flight, per-agent latency.
Cost and Latency
Wall time per verified agent, completion rate and cost per verified task.
One mesh. Every machine's model and compute in reach.
Every device running mimOE becomes part of one mesh. Run a model on one machine and use it from mimOE Studio on any other, as if it were local.
- Every machine running mimOE is a node.
- Add the machines you already have.
- Put each model where it fits.
- Use one big GPU machine from all the others.
Use another machine's model and compute.
The model on the GPU box down the hall appears in Studio like a local one.
Any PC, Mac or iOS device.
If it runs mimOE, it's in the mesh.
Discovery is automatic.
Instances find each other over link-local mesh discovery. No IPs, no setup.
Stable node IDs.
Addresses change. Node IDs don't, so Studio always finds the right machine.
Attach a tool once. Every harness has it.
Attach an MCP server to mimOE once and it is available in Claude Code, Codex and Cursor. Each client connects with its own key, limited to what that key may access, and you can revoke any key from one place.
Four moves, and it is set up.
The install script starts the engine. You only need to set up each server once. Then, for each client, add its unique key and a single line of configuration.
The models come from wherever you already run them.
You can also connect a local OpenAI-compatible server you already run, or add a cloud provider using an API key. Their models appear in Active models, under Models, and work like any other model.
- ollama
localhost:11434qwen3.6:35bOn-device servers - ovms
localhost:8000OpenVINO/Qwen3-8B-int4-ovOn-device servers - lemonade
localhost:8001Qwen3-4B-GGUFOn-device servers - groq
api.groq.com · keyedllama-3.3-70b-versatileCloud, when you choose it