One 59MB binary. Your whole AI operating layer.
The engine underneath every LibraOS digital employee is a single Go binary that runs inside your own environment — air-gapped if you need it. Multi-agent orchestration, grounded retrieval, a 3-tier firewall, and 100+ models behind one OpenAI-compatible endpoint. No vendor cloud. No data egress.
The cascade: the LLM runs only when it's needed.
Most requests never reach a frontier model. A three-tier cascade resolves the routine in milliseconds and escalates only what genuinely needs reasoning — so latency and cost track the work, not the model.
Tier 1 — deterministic. Regex, cache, and policy resolve known-shape requests instantly, with zero token spend.
Tier 2 — small-model routing. A fast local model classifies, retrieves, and answers the grounded majority.
Tier 3 — frontier reasoning. The heavy model runs only when the first two tiers defer — the exception, not the default.
Four systems, one binary.
The engine ships as a single artifact. Every digital employee, every function, every pack runs on the same four systems — nothing to assemble, nothing that phones home.
Cascade routing
Three-tier orchestration plans the work and routes it. Most requests resolve in milliseconds; the LLM runs only when the cascade defers to it.
Grounded retrieval
Graph + vector + keyword retrieval with cross-session memory. Every answer traces back to your source documents — grounding, not guessing.
3-tier screening
Regex → ML risk scoring → LLM judge, on both inbound and outbound traffic, for prompt injection, jailbreaks, and PII. Utility holds at 90.7% with defenses on.
100+ models
Frontier and open-source models behind one OpenAI-compatible endpoint with intelligent failover. Point your existing clients at it — zero lock-in.
Numbers that survive diligence.
Controlled same-model A/B, every figure certified across three runs. Full methodology and per-run data available on request.
+18.9 pts
GAIA (public agentic benchmark, all levels). Same brain, with vs. without the engine. Replicates on-prem: +17.2 apples-to-apples on the shipped local brain.
256 / node
Grounded digital employees per 8×H100 node. 98.3% grounded success across ×3 runs — we measure grounded success, not HTTP 200.
−35%
Attack success rate on AgentDojo — 26.8% → 17.3% with defenses on, while benign utility holds at 90.7%.
Deploy where your data already lives.
One binary, your infrastructure. About two weeks to production — and your existing OpenAI-compatible clients keep working unchanged.
Built with Go. Runs anywhere. Compatibility identifiers from earlier releases (env vars, client IDs, image tags) are honored indefinitely.
Put the engine to work.
A 30-minute call with our engineering team — we'll run it against your documents, your compliance rules, and your security requirements.
Read the integration guide, or open the console to point an existing client at the endpoint.