KRNL OS is the planned Linux-forked custom kernel direction: a benign machine workspace where policy, memory, models, workers, and agents become visible system services. The OS itself is not shipping yet; its foundation is being proven in the Harness.
The goal is a transparent local boot chain where each intelligent service has a bounded role, a health state, and an observable relationship to the operator.
The planned OS turns today's Harness boundaries into first-class machine services rather than hiding them inside a monolithic assistant.
Coordinates sessions, events, heartbeat, agents, and durable task state.
Resolves capability requests and keeps human authority in consequential workflows.
Builds retrieval-backed context and routes proposed knowledge through approval and quality scoring.
Selects local or hosted intelligence according to policy and machine capacity.
Executes skills through isolated language-specific runtimes.
Records a BLAKE3 hash chain with quantum-entropy-enhanced salts and provider provenance.
MAGIC — the Mixed Authenticated Quantum Intelligence Core — gives Karnelian a provider-aware entropy pipeline. It can use Quantinuum H2 through pytket, Qiskit-based sampling, or a secure operating-system CSPRNG fallback.
The repository includes quantum-category Python worker skills for random sampling and optimization research. The fallback remains cryptographically safe when quantum providers are unavailable.
Route between local and hosted intelligence based on policy, privacy, and machine capacity.
Apply deny-by-default capabilities and human approvals across every agentic process.
Assemble relevant context across sessions and tools while keeping ownership local.
Run bounded tasks through Node.js, Python, WebAssembly, and trusted native execution.
Verify privileged event history with BLAKE3 chaining and optional chain anchoring.
Coordinate heartbeat, agents, events, task scheduling, and persistence.
