Your coding agent.
Its own inference engine.
Open source, verifiable.

Built for Tribunus Desktop — the AI coding agent. Powered by Tribunus Compute — the portable inference engine with compile-time architecture.

Download Desktop Explore Compute

Tribunus Desktop agent

Your AI coding agent. Runs locally. Plugin system. Multi-agent. Any model, any backend. Electron desktop app for macOS, Linux, Windows.

  • Agent system with tool execution
  • Project-aware sessions
  • Any LLM backend (local or remote)
  • Plugin system for custom tools
Learn More

Tribunus Compute engine

The inference engine inside Tribunus Desktop. Compile-time architecture — every decision frozen before runtime. Numerical oracle validates every kernel. Multi-backend from day one.

  • No guessing: compile-time candidate generation + 6-check admission
  • Numerical oracle: Apple Silicon FP32 as reference
  • Multi-backend: CUDA, Metal, Vulkan, ROCm, oneDNN, Level Zero, TT-NN
  • 6 runtime pipelines: token intake → prefill → decode → KV → speculation → output
Learn More

For Engineers

Architecture decision records that define the Tribunus Compute engine — each one generated, verified, and source-backed.

▶ $ ADR 0037: Backend Realization Contract ✓ verified

Defines the formal contract each hardware backend must satisfy to be admitted into the Tribunus Compute runtime. Covers kernel registration, memory layout, stream synchronization, and the 6-check admission pipeline that gates every backend at compile time.

Full ADR Source
▶ $ ADR 0038: Numerical Governance + Autotuning ✓ verified

Establishes the numerical oracle — Apple Silicon FP32 as the golden reference — and the autotuning framework that validates every kernel against it. No kernel ships without passing the oracle's accuracy and performance gates.

Full ADR Source
▶ $ ADR 0039: Datacenter Control Plane ✓ verified

Specifies the control plane architecture for multi-node inference deployments: node discovery, health monitoring, load distribution, and the contract between the control plane and the per-node Compute runtime. Enables scale-out without sacrificing compile-time guarantees.

Full ADR Source
▶ $ ADR 0040: Runtime Inference Pipelines ✓ verified

Defines the six runtime pipelines — token intake, prefill, decode, KV cache management, speculative decoding, and output streaming — their scheduling guarantees, memory budgets, and the handoff contracts between adjacent stages. Each pipeline is a separate phase in the compile-time planner.

Full ADR Source
▶ $ ADR 0041: Server UX Strategy ✓ verified

Defines the user-facing inference server interface: REST API surface, streaming protocol (SSE + server-sent events), health and metrics endpoints, configuration model, and the development-server mode that mirrors production behavior without requiring a full cluster.

Full ADR Source