Define a goal. Nova breaks it into subtasks, executes them through a coordinated agent pipeline, evaluates progress, re-plans, and completes — on your hardware, under your control.
Define a goal. Nova breaks it into subtasks, executes autonomously, re-plans as needed.
Learns your preferences, customizes itself, updates its own configuration over time.
Runs on your hardware. Your data never leaves. Credentialed actions land in an HMAC-chained audit log; mutating actions require explicit user consent.
Continuous batching, concurrent pipelines, multiple inference backends. No bottleneck.
Agent Pipeline
Every task runs through specialized agents with built-in safety rails. No shortcuts.
Curates relevant code, docs, and task history
Produces the actual output — code, config, or answer
Checks for prompt injection, PII, credential leaks, spec drift
Pass, needs refactor, or reject — loops back up to 3 times
Creates decision record and escalates to human on rejection
Post-Pipeline
parallel, non-blockingCapabilities
From local inference to cloud routing, from IDE integration to self-configuration.
vLLM, Ollama, or cloud — select from the UI. Nova handles container lifecycle, health monitoring, and graceful backend switching.
Consent gate for mutating tool calls. Encrypted credential vault. Hash-chained audit log. GitHub provider with READ/PROPOSE/MUTATE/SETUP tiers.
Cortex watches GitHub webhooks; failing CI dispatches a goal that proposes a fix PR. Per-cycle cost budget; nothing merges without your approval.
Optional screenpipe-bridge ingests your screen activity into Nova's memory with a privacy denylist (apps, URL patterns, window titles) and pause-without-disconnect.
Graph-based associative memory with spreading activation, sleep-cycle consolidation, neural reranking, and source provenance. Bounded query plans verified by 200+ unit tests.
Plug in any MCP server: GitHub, Slack, Sentry, Playwright, Docker, and more.
Anthropic, OpenAI, Ollama, Groq, Gemini, Cerebras, OpenRouter, plus subscription-based Claude/ChatGPT.
Auto-detects GPU hardware, recommends inference backends, manages containers. Supports remote GPU over LAN with Wake-on-LAN.
Backup/restore, factory reset, service health monitoring via dedicated sidecar service. Docker SDK gated behind socket-proxy for blast-radius reduction.
OpenAI-compatible endpoint works with Cursor, Continue.dev, Aider, and any OpenAI-API client.
Talk to Nova hands-free with Gemini-style conversation mode. Barge-in interruption, live transcription, auto-listen between turns.
Get Started
The setup wizard detects your hardware, configures providers, and starts everything.
Remote GPU
Got a separate machine with a GPU? Run the remote setup script on it, then re-run
./install
and choose "Remote GPU".