← Blog
· OLAV Team

OLAV v0.10.0 — Open Source Launch

OLAV is now publicly available under BSL-1.1. Here's what's inside the platform, how the architecture works, and what comes next.

release architecture open-source

We’re open

After months of internal development, OLAV v0.10.0 is now publicly available under BSL-1.1.

The full platform core — CLI, agent runtime, audit chain, API registry, and Creator Agent — is open and free to self-host.

How it works

OLAV sits between your team and your infrastructure APIs. Every operation flows through four governance layers before execution:

graph TD
    U[User / CLI] --> AAA[Layer 0 — AAA\nAuthN · AuthZ · Audit]
    AAA --> MW[Layer 1 — Middleware\nCache · Rate-limit · Retry]
    MW --> SB[Layer 2 — Sandbox\nSchema validation · Dry-run]
    SB --> OUT[Layer 3 — Output\nNormalise · Sign · Emit]
    OUT --> API[Downstream API]

    style AAA fill:#7f1d1d,color:#fca5a5,stroke:#991b1b
    style MW fill:#78350f,color:#fcd34d,stroke:#92400e
    style SB fill:#14532d,color:#86efac,stroke:#166534
    style OUT fill:#1e3a5f,color:#93c5fd,stroke:#1d4ed8

No operation bypasses this chain. The audit record is written before execution and is immutable.

Creator Agent — registering new API skills

When you point OLAV at a new API, the Creator Agent walks through six deterministic steps to produce a deployable skill:

sequenceDiagram
    actor Dev as Developer
    participant CA as Creator Agent
    participant DB as DuckDB Registry
    participant API as Target API

    Dev->>CA: Register skill (OpenAPI schema)
    CA->>CA: Parse & validate schema
    CA->>API: Probe endpoints (dry-run)
    CA->>DB: Store skill + constraints
    CA->>CA: Generate semantic embedding
    DB-->>Dev: Skill ready ✓

The result is a callable, audited, semantically-indexed skill — no custom glue code required.

Self-evolution loop

OLAV learns from failures. When an agent operation hits a constraint violation, the failure pattern is absorbed as a new rule and immediately applied to future queries:

flowchart LR
    Q[Query] --> M{Match in\nsemantic cache?}
    M -- Hit --> R[Return cached result]
    M -- Miss --> A[Agent execution]
    A --> V{Constraint\ncheck}
    V -- Pass --> O[Execute + Audit]
    V -- Fail --> E[Extract failure pattern]
    E --> C[Emit new constraint rule]
    C --> M

Over time, the constraint ruleset grows, and repeated failure patterns vanish.

Performance

The semantic cache gives dramatic speedups for repeated or paraphrased queries:

Query typeLatency
Exact cache hit< 1 ms
Semantic cache hit~5 ms
Full agent execution800 ms – 2 s
Cache hit vs full exec2000× faster

What’s next

Read the full docs → · View on GitHub →