Features and limitations¶
What you can do¶
-
Download a fixed-version v1.0.0 USB installer with a public checksum and build evidence. Future release delivery includes an automatic exact-version installer build and full public-download verification; the product tag can appear before that completion. Media integrity is not physical installation or hardware acceptance.
-
Choose fixed released versions under System → Settings → Updates, review release notes and original license dates, then check/apply the same host/Flux revision. Prereleases are opt-in; unavailable legacy releases/offline catalogs never fall back to main. Development branches remain explicit advanced options. See software versions.
-
Install new Ubuntu 26.04 appliances through the normal host bootstrap with a reviewed, enforcing containerd AppArmor compatibility profile loaded before K3s. No global AppArmor disablement or kernel downgrade is used; existing unowned clusters are not silently changed. See host bootstrap boundaries. Fresh VM/physical installation acceptance remains separate from local checks.
- Run local models with Ollama or vLLM on eligible CPU/GPU targets.
- Create and edit Ollama, vLLM and experimental vLLM-Omni in one local-model configuration form, with shared runtime/CPU controls and engine-specific source, hardware and memory options.
- Detect whether a local model provides Chat or Embedding from its metadata; choose its task manually only when it is unknown.
- Use NVIDIA GPUs on one node, including different GPU models and capacities, to split one model or run independent copies behind one LiteLLM model name, with explicit per-card allocation through NVIDIA DRA.
- Select physical NVIDIA cards from one checkbox list and see consistent total VRAM estimates and breakdowns for a split model. Validate model names before creating local or external models.
- Review separate GPU-model estimates for serving RAM, startup range and a conservative host budget. New multi-GPU/offloading defaults are model-dependent; saved and manually edited budgets remain unchanged. Estimates are not fit guarantees.
- Experiment with Realtime by selecting the vLLM-Omni profile in that form.
- Connect external model providers and use a common LiteLLM API.
- Give every application instance its own LiteLLM inference key, with automatic creation, suspension, rotation and cleanup; application workloads receive no LiteLLM administrator key.
- Preserve known tool, image-input and reasoning capabilities through the shared model catalog and native agent configurations, including explicit unsupported capabilities.
- Create applications from the service catalog and grant access to specific users or groups.
- Keep separate OpenClaw and KubeOpenCode model choices while sharing the common model catalog. Catalog updates preserve each instance's selection.
- Open unmodified upstream OpenCode instances through
<name>.oc.<domain>; their HTTPS address, SSO callback and local certificate stay aligned without a custom OpenCode image or a per-browser server setting. - Choose separate initial chat and embedding models for AnythingLLM in Services → Configure, or use the catalog defaults. Retain its own model preferences across restarts. Existing installations need the one-time settings migration.
- Create Hermes instances with an explicit LiteLLM model and open the current Hermes dashboard through the common SSO access controls, with persistent sessions and workspaces.
- Create Paperclip instances for companies, tasks and OpenCode agents, with isolated Kubernetes workspaces, LiteLLM models and the common SSO access controls. The pinned onboarding offers the instance's existing Magic Stick model connection without a cloud subscription. The current sandbox runtime requires AMD64.
- Run the optional Pi Coding Agent from Services, with a browser terminal, model selection, SSO access controls, and persistent workspaces and sessions.
- Share running local chat models through the opt-in Private Mesh module.
- Manage local users, GPU allocation, host networking, model cache and Ubuntu updates.
- Plug in Ethernet after Wi-Fi setup on a new installation: an unconfigured wired port can get its address via DHCP. This does not by itself verify access to the dashboard after switching links.
- Inspect System → System Status through separate Overview, GPU Operators, Flux and Network & routes tabs; administrators also find Model cache and Regression tests there.
- Run the existing P0 suite from System → System Status → Regression tests on a dedicated, registered test Appliance without a license; see progress, Passed/Failed/Blocked results and download sanitized reports. Filtered diagnostics retain the first failed baseline check and its observation history, including safe Kubernetes events and container restart/exit states; raw event messages and optional bounded controller logs remain private. Run all phases by default or select one or more, such as Phase 3 only; progress and reports retain that selection without crediting unselected phases. Phase 4 also offers fixed diagnostic subsets for slot races, reload and exact restoration, individually or together. These retain automatic preparation and cleanup and never certify the whole phase. Choose Non-destructive only (default) for isolated tests and live read-only inventory, or Include destructive tests for model/service/configuration workflows. The limited mode creates only test infrastructure and a temporary viewer account; excluded live steps are named, not counted as passed. Product-service/application live workflows, including Hermes, service credentials and app access, are currently deferred to separate regression. The explicit separate Services selection tests catalog-derived installation and removal, access with SSO and supported public access, including Pi Coding Agent. Model integration remains deferred; Hermes remains separate. Ordinary runs keep their exclusions. Isolated checks and core identity/license/Kubernetes/model/GPU coverage stay active; reports name exclusions without claiming service/application acceptance. Fixed diagnostic selections can repeat existing external-provider, GPU-verification, two-card independent-copy Ollama, two-/four-card shared split and sharing scenarios without claiming that the complete phase was tested. Clean up the previous run separately, including retained setup access when its lock is already free; disabled actions explain their blocking condition. Distinguish missing prerequisites from checks blocked by an earlier failure, with the original blocking check and trace when recorded. Automatic preparation failures show their setup stage and next action. All checks includes isolated and live evidence, including runner and cleanup errors; Executable scenarios has separate, initially closed details. Live evidence stays separately labelled. Missing final totals are not shown as zero failures; 100% progress means execution finished, not passed. Expand Cleanup observations for the original cause, remaining resources and observed wait budget. Sanitized traces retain useful reviewed step names, including shared helpers; credentials and raw responses stay excluded. Kubernetes cleanup details distinguish revision conflicts, access denial and unavailable responses; safe conflict retries retain ownership checks. A report-download export failure retains completed results and does not by itself require cleanup or repeating the tests. Each full start cleans up the previous run and begins again with a fresh test ID. Non-destructive starts require any previous-run cleanup to be completed separately. Clean up previous test run is also available for diagnosis; cleanup verifies the retained resource identities and revisions. Interrupted Kubernetes executions require exact termination evidence as well; ended Linux zombies are distinguished from live workers, and original runner failures remain visible after cleanup. Test completion, verified resource cleanup and test-account removal are shown separately; an old report without cleanup evidence is not labelled clean. Host-disconnecting drills still require an external runner.
- Use the browser dashboard or authenticated CLI/TUI; advanced users can manage runtime resources.
- Operate the appliance at its own monitor and keyboard without logging in: the terminal dashboard signs in automatically, also without a network connection. Administrators can lower its access or switch it off, and any user can switch to their own account. Automatic sign-in.
- Optionally select Remember me on the local login page to stay signed in across browser restarts: seven days idle, thirty days maximum. Unchecked sessions keep the thirty-minute idle limit. Signing in.
- Browser SSO cookies stay on each dashboard/service hostname instead of growing one appliance-wide cookie header. Services share the central sign-in; global logout invalidates the session across those hosts. Missing gateway protection denies access rather than making an application public. Parallel page files share only a currently running host-readiness check; each request still passes its own authentication and access controls. Paperclip keeps browser sign-in separate from its native agent credentials; the central browser token is not sent to its agent-token verifier. Revisiting a logged-out application in the browser starts sign-in again; a permission denial or unavailable identity service does not restart login. Existing SSO access can migrate even with a stopped application model, while invalid instances remain blocked and no new workload is provisioned.
What is not automatic¶
A detected GPU is not proof that every engine or model will run. The dashboard checks engine capabilities, driver readiness, allocation slots and memory data. See the compatibility matrix.
Ordinary vLLM runtimes use upstream 0.31.0, pinned separately for CPU, NVIDIA, AMD ROCm and Intel XPU. A supported architecture does not guarantee a specific quantization or community checkpoint works on your GPU. Experimental vLLM-Omni retains its separate version and acceptance boundaries.
Ollama uses digest-pinned 0.40.3 standard and AMD ROCm images. Saved registry references and model controls are retained; upgrading an image does not certify every model or GPU. See runtime upgrade checks.
GPU sharing permits more workloads to use a device; it does not multiply its physical memory or provide isolated VRAM limits. Memory estimates do not prove that a particular context length will fit. Downloading large checkpoints may also need substantial disk space beyond their final size.
Multi-GPU placement requires NVIDIA cards on the same node and DRA with known physical capacities. A split model divides its budget evenly: the smallest selected card's unreserved budget bounds every share and therefore the total. Each independent copy needs its own full memory budget. Engines, model formats and quantizations must work on every selected card; mixed-vendor and multi-node groups are unsupported. Physical multi-GPU inference and throughput still require acceptance on the intended hardware.
vLLM-Omni has its own runtime boundaries. An OpenAI-compatible
chat endpoint does not automatically provide /v1/realtime or audio output.
Pi requires GitHub access when creating or recreating its runtime. Each instance accepts one browser connection at a time, and its users share files and sessions.
Hermes 0.21.5 requires a chat model with at least 64,000 configured context tokens. Unknown context metadata needs an administrator to declare the actual supported limit before the instance can be installed or upgraded.
There is no appliance-wide one-click backup/restore or supported factory-reset button. Backup and recovery requires a planned, separately tested procedure. A Flux update does not upgrade the Ubuntu release.
Access and licensing¶
Local identity, model runtimes, GPU management, Resource Sharing and Private Mesh do not require a license file. Federated SSO requires Free Registered or Commercial activation. Technical activation and legal permission for production use are different questions; LICENSING.md and LICENSE are the authoritative terms.
See license management for the dashboard workflow.