The Agencie Lab
A frontier AI lab, on your own hardware.
Your choice of models, hosted privately. Autonomous agentic pipelines, full coding, vision, documents, slide decks, voice and image generation, and a browser for agents, through one governed endpoint, with a budget, a check and an audit record for every request, and runtime security from SILO.RED. It is what we use internally, now offered privately to the right partners.
Overview
A frontier provider in miniature.
Coding agents are now how good teams ship. The catch is that they run on someone else’s models, read your code on someone else’s servers, and bill by the token with nothing to say what the money bought.
The Agencie Lab is two machines and three licensed modules. The Docker host runs AOS, the gateway and console, and Code Easy, which knows your repositories. The models host runs open models on a GPU, a DGX Spark included. Your people keep the tools they like: CloudCode in the terminal, VS Code with our extension, Continue or Cline, or any OpenAI-compatible client.
We supply the hardware, install the stack, register the repositories, set the budgets with you and hand over with the runbook. Your frontier subscriptions or keys go in on the day, and nothing of yours passes through us afterwards. The licence pays for the platform; your usage never leaves your lab.
What it does
Governance in the platform, not in a policy document.
One door for every tool
CloudCode in the terminal, VS Code with Continue or Cline, Code Easy, scripts and CI all call one gateway. It chooses the model, offers the tools, enforces the budget and keeps the record.
Local first, frontier by choice
Routine and standard work runs on open models on your own GPU. The hardest tasks go to a frontier agent as a whole, checked job, and only when you switch that on. Switch it off and nothing leaves the building.
Checked, as a provider checks itself
A coding turn ends only when the project’s own tests have run. A reviewer reads the change against the brief. Trust is earned from verified outcomes, never granted.
Nothing runs unaccounted
Every request reserves money before it starts and settles after. Limits per day, month and client; alerts at 80% and 100%; a monthly statement by team, repository and model for charge-back.
Autonomous builds
Code Easy knows every registered repository. AutoCode turns a specification into a build: one engine writes, another validates, a stuck task escalates. Close to a hundred specifications have run on our own products.
More than code
PDFs, Word, Excel and PowerPoint files are read, and made: documents, spreadsheets, PDFs and slide decks come back as links. Screenshots and photographs go to a vision model; pictures and voice-overs are made on the lab’s own servers, and short video with sound on request. All of it on your hardware, budgeted and recorded.
A browser for agents
Agents open web pages and work them as a person would: read, click, fill in a form, follow a link. Each page comes back as text the lab’s own models can read, so testing your sites becomes part of a build. Kept to the lab whenever you choose.
A mode for the work in hand
Isolated: no frontier model is sent anything. Max mode: the idle stacks stop and AOS has the machine. Studio mode: the Spark swaps its reasoning model for video and images. Each is a switch on the console, applied by the host and reported step by step.
The console, from anywhere
A live map of every model, tool, agent and host; spend to the penny; CloudCode itself in a browser tab, on a desk or a phone. One sign-in, roles for who may change what, and every key in an encrypted vault.
In use
Watch it think, from a desk or a phone.





Who it is for
Teams that want the frontier, on their terms.
- Engineering teams spending thousands a month on metered coding agents with no governance to show for it.
- Regulated firms whose policy says code never reaches a frontier provider.
- Groups with several teams or sites who need spend charged back by team and repository.
- Anyone who wants the frontier way of working without handing over their code or their budget.
Ways to buy
Priced per lab, installed and supported.
Hardware passes through at cost. A quote follows a conversation.
- The mini lab: the host, the stack installed, the governance on, a day of handover.
- With the DGX Spark: a 120B reasoning model on the complex tier, on premises.
- A hardware-ready kit: built and tested here, shipped, plugged in.
- Deployed on your own machines or private cloud: software, installation and support.
- Operated for you: your hardware, our engineers watching the console.
- Institutional: your accelerators, isolation from day one, and the evidence for an auditor.
- Video generation on the same hardware, for teams that need it: ask us.