Cluster operation
FICC
Federated Industrial Cluster Commander.
Cluster operations, local inference and connected services.
FICC brings machine inventory, managed jobs, files, terminals and coding agents into one workspace. The local controller connects through approved OpenSSH profiles and retains operation history in its own database.
Observe resources, preview commands and follow their output. Browse registered roots, verify transfers and keep terminal sessions grouped by machine. Named agent runs record explicit message recipients and delivery receipts.
New: 0.3.0.dev1 nightly
The first 0.3.0 nightly brings model management, verified weight storage and managed inference to FICC. Choose an approved contributor, reserve its GPUs and start a service with llama.cpp, Ollama or optional vLLM. A model can span GPUs on one machine; separate machines can run replicas. Cross-host tensor parallelism is not supported.
HyperNet federation connects paired clusters to selected inference services, explicitly granted lifecycle operations and resumable, verified model-file transfers. HyperNet Core and Suite 0.9.0 remain separate installations, with reusable networking capabilities beyond FICC.
Individuals can operate models across a workstation or home lab. Organisations can share selected services between approved teams or sites, reuse model weights and retain operational records. Managed inference is qualified on Linux with NVIDIA GPUs; engine images and models are separate downloads.
This is a nightly prerelease. The stable release remains 0.2.6r4-final. Back up controller state before upgrading; database rollback requires the matching pre-upgrade backup.
Local control, explicit operations
Each machine has a pinned SSH identity. Jobs and terminals use the authority of the remote account. The console shows uncertain outcomes and retains the records needed for recovery.
Stable: 0.2.6r4-final
The final 0.2.6 revision discovers installed coding agents on enrolled Linux and Mac machines. Choose a machine, select its agent and review the launch. Refresh updates the inventory; Back preserves the choices in an unsubmitted preview. Discovery leaves installation, provider sign-in and model selection with the operator.
Apple Silicon GPU observations show available utilization and used unified memory without inventing dedicated capacity. AMD integrated GPU readings use driver-reported VRAM. Reported hardware observations cover Apple M5 Pro and AMD Raphael; discrete AMD cards, including the RX 590, remain unqualified. Apple and AMD devices remain observation-only.
Agent discovery and launch → GPU readings and qualification → Security model →
Appearance and themes
Version 0.2.6 adds 26 themes, including the original Classic Light and a Classic Dark palette that matches AOTX PRISM. Custom JSON themes can change colors, gradients and the logo. A flat-material option removes gradients and the background grid. Responsive navigation, aligned controls and floating-panel restoration improve workspace use.
Read the appearance and custom theme guide →
Connected workflows
Version 0.2.5 adds project access, explicitly configured HTTPS remote access, contributor workload queues and dataset pipelines. Registered data sources feed verified dataset versions; completed contributor output can be published with retained lineage. Optional local inspection, encrypted secret references and separately administered audit delivery add deployment controls.
Contributor runtimes, external services and access grants need explicit setup. Dataset traffic currently relays through the controller, and local inspection records coverage rather than certifying arbitrary content as safe. Linux packages include a signed checksum manifest; verify it using the independently trusted publisher key before installation.
The FICC website
The dedicated site explains the capabilities and provides Linux packages, native Mac installation guidance and searchable manuals. Its documentation covers operation, access, retained state, backup and recovery.