Deskarden’s deployment model keeps workspace storage, agent orchestration and model inference within a boundary you choose.
The browser communicates with your workspace host. Agent workers retrieve selected project context and call a local model endpoint. Outputs return to the project for review.
No additional hosted model provider is required for local inference. Downloading model weights, installing updates, web research and enabled connectors can still require external network access.
Explore data boundaries →There is no single useful minimum for every model and workload.
| Resource | What to account for |
|---|---|
| Model memory | Weights, quantization, context length and active requests. Follow the chosen model and runtime’s requirements; leave room for the workspace and operating system. |
| Compute | CPU inference can suit smaller models and patient workloads. Supported GPU acceleration can increase throughput, depending on model and runtime. |
| Storage | Source files, model downloads, prepared context, accepted outputs and backups. Fast local storage helps file preparation and model loading. |
| Concurrency | Start with a modest worker limit. Watch memory and queue time before increasing simultaneous runs. |
| Availability | A sleeping personal computer cannot serve a shared workspace. Plan service startup, maintenance windows and recovery for a team host. |
| Network | HTTPS for remote browser access. Keep model endpoints private and enable outbound connectors deliberately. |
Choose a model for its task quality, resource needs and license. “Open weights” describes access to model parameters; usage rights still depend on the license.
A local runtime handles model loading and inference. The workspace’s model configuration describes the endpoint, model identifier, context budget and worker limits.
Runtime-specific setup belongs to the runtime’s own documentation. Useful starting points are Ollama’s local runtime guide ↗ and llama.cpp’s project documentation ↗.
Some runtimes also offer cloud features. Confirm which model endpoint a workspace uses and disable unwanted external options.
Long documents, structured output and tool use place different demands on a model.
Limit active workers and context size. Test a representative project before growing the workload.
New deployments are arranged directly. Tell us about your work, your host and the people who need access; software and support terms are agreed before setup.
Your computer, hosting, electricity and any model-license costs remain separate. This website does not take payments.
Discuss a deployment →Sign in with the account email assigned to your workspace. A membership invitation is issued by its administrator.
Bring your projects, models and workflows into one private workspace.