olav 0.11.0

OLAV — AI-native platform for autonomous operations with domain-aware agents, API-as-Tool, and Agent Harness

OLAV 🐺

Online Analytical Vertex for Agentic Operations
AI-native platform for autonomous infrastructure operations.

中文文档

Control your infrastructure with natural language. Connect any API, deploy specialized AI agents, orchestrate operational workflows — without writing SQL or memorizing CLI flags.

olav "how many devices are online?"
olav "which BGP neighbors are down?"
olav --agent core "run a health check on all endpoints"

Quick Start | 中文

---

Why OLAV?

API-as-Tool — Any API Becomes an Agent Tool in One Command

No code, no MCP servers. Register an OpenAPI service once, query it in natural language forever:

olav registry register http://netbox.example.com/api/schema/
olav "how many devices are in rack A1?"   # works immediately

The Creator Agent can also generate editable Python @tool functions from any OpenAPI schema for deep customization.

Agent Harness — Four-Layer Execution Governance

Every Agent decision passes through a mandatory execution control layer:

Layer 0: AAA        Token/LDAP/OIDC auth → RBAC → full audit trail
Layer 1: Middleware  HITL dangerous-command interception + memory injection
Layer 2: Sandbox    Pre-scan → DuckDB read-only enforcement → network namespace isolation
Layer 3: Output     Credential auto-redaction + SSE JSON encoding + HttpOnly cookies

Hard constraints (DuckDB read-only) cannot be bypassed. Soft constraints (injection scanning, network isolation) are configurable.

Self-Improving Loop — Agents Get Better Over Time

Use OLAV → audit log captures every tool call and error
    → /trace-review extracts failure constraints → writes to LanceDB memory
    → future runs recall constraints before acting → known pitfalls avoided

Data-driven improvement, not manual prompt tuning.

Multi-Layer Cache — Measured 2000x+ Speedup

Tier-0: SemanticCache (LanceDB vector similarity, ~10ms)
Tier-1: LLM SQLiteCache (exact prompt match, <1ms, 0 tokens)
Tier-2: LLM API call (real network request)

Per-user isolation. Anthropic models automatically use prompt caching (system prompt billed once).

Federated Specialist Agents

Not one omniscient Agent — multiple specialists collaborating:

• Semantic router dispatches queries to the best-fit Agent based on route_keywords
• Each Agent holds only its own tools (least privilege)
• Per-Agent model assignment: cheap models for frequent queries, powerful models for complex analysis — 40-70% token cost reduction
• Hot-swappable Skills: olav skill install extends capabilities instantly

AAA — Authentication, Authorization, Audit

• Auth: none / token / LDAP / AD / OIDC; tokens stored as salted SHA256 hashes
• RBAC: 3 roles (admin/user/readonly) × 5 actions, fine-grained per Agent/Skill
• Audit: Every operation recorded to DuckDB (4 tables), SHA256 tamper-proofing (NIST AU-9), credentials auto-redacted

Full-Stack Observability — Audit as Data Asset

Audit logs are not just compliance — they're a continuously growing data asset:

• Query any dimension with DuckDB SQL (token usage, cache hit rate, tool call frequency, error patterns)
• Multi-user concurrent-safe (DuckDB atomic writes, each record tagged with user_id)

---

Quick Start

=== "pip install (recommended)" ```bash # 1. Install pip install olav

# 2. Initialize project
olav init                                   # creates .olav/ with config skeleton

# 3. Configure API key (edit the generated file)
#    Set shared.api_key in .olav/config/api.json
#    Or use environment variables:
export OLAV_LLM_API_KEY="sk-..."           # your LLM provider API key

# 4. Connect a service and start querying
olav registry register http://netbox.example.com/api/schema/
olav "how many devices are in rack A1?"

# Or install a community skill
olav skill install https://github.com/olav-ai/skill-netbox
olav "list all sites in Europe"
```

=== "From source (development)" ```bash # 1. Clone and install git clone https://github.com/olav-ai/olav.git && cd olav uv sync

# 2. Initialize and configure
uv run olav init                           # creates .olav/ with config skeleton
# Edit .olav/config/api.json → set shared.api_key
# Or: export OLAV_LLM_API_KEY="sk-..."
uv run olav registry register http://netbox.example.com/api/schema/
uv run olav "how many devices are in rack A1?"
```

Other ways to use

olav                            # interactive TUI (multi-turn conversations)
olav service web start          # web UI at http://localhost:2280
olav --agent core "run: df -h"  # execute shell commands via Core Agent

---

Three Interfaces

Interface	Command	Best For
CLI	olav "your query"	Scripting, one-off queries, CI/CD
TUI	olav	Multi-turn conversations, exploration
Web UI	olav service web start	Team sharing, browser access

---

Architecture

User Query
    ↓
Semantic Router → selects best Agent (or --agent override)
    ↓
Agent Harness → AAA → Middleware → Sandbox
    ↓
LLM + Tool Loop (calls tools, gets results, synthesizes response)
    ↓
Response + Audit Log (automatic, every run)

Tech Stack: LangChain + LangGraph + DeepAgents + DuckDB + LanceDB + FastAPI

---

Repository Structure

src/olav/          ← OLAV core platform (this repo)
src/README_ZH.md   ← 中文文档
src/olav_logo.png  ← Logo
.olav/workspace/   ← Platform agent definitions (config/ + core/)
.olav/config/      ← Runtime config (gitignored — contains API keys)
.olav/databases/   ← Runtime data (gitignored — audit logs, domain data)

---

Documentation

Documentation: docs.olavai.com

Website: olavai.com

---

License

BSL-1.1 — Business Source License 1.1