Assess the work
Start with the use case, recommended model, and expected team demand.
Private AI on hardware your organisation owns. Start a focused Claude Code pilot on a suitable Apple Silicon Mac. CaseDesk guides setup, owner approval, client access and status while the model runs on hardware you control. Establish the value of a real coding workflow before expanding to your team.
No public IP, inbound firewall rule, or Kubernetes required.
Your local runtime stays on hardware your organisation controls.
Approved model. Health visibility. Controlled maintenance.
The local runtime runs the model. CaseDesk governs access and approved changes around it. Local runtimes are excellent tools for an individual developer; when a team depends on one machine, the operating questions change: which model is approved, who owns maintenance, whether the device is ready, and how internal tools connect without opening the runtime to the internet.
CaseDesk does not take over the hardware. It provides the approval and connection layer that lets you evaluate an approved coding workflow without exposing the local runtime directly.
Start with the use case, recommended model, and expected team demand.
An owner authorises a Device Agent on the host. The first report is read-only.
Review capacity, model identity, runtime changes, and operating limits before anything changes locally.
Keep the runtime private while the connection reports health and maintenance state.
CaseDesk does not buy, operate, or pay for your hardware. It gives the local runtime a governed connection, a visible owner approval path, and a route to repeatable operations as demand grows.
Give internal developer tools an approved connection to a local model without exposing the host.
Use hardware you already own when utilisation, device support, and availability expectations are understood.
Qwen3 4B has been retired from new Claude Code and OpenCode coding trials. Existing approved connections remain available while their replacement is prepared. Devstral Small 2 is available for supervised trials on compatible Apple Silicon Macs under an owner-approved plan. Each installation must pass its own capacity and coding-workflow checks. Review generated changes and run tests before using them. This trial is not qualified for unsupervised coding.
The reviewed Mac plan uses a 65,536-token context and one concurrent request. Keep the host awake and connected. Validate memory, storage, response time and your project before depending on it for daily work; simultaneous team use and full Mac reboot recovery need separate testing.
Inference runs on your Mac. Claude Code sends requests through CaseDesk, and responses return through its gateway. This is not an offline or air-gapped service. Include that path in your organisation’s security review.
Talk through your workload, device ownership, operating boundary, and the path to a governed private AI service.
Talk to the CaseDesk team