olav 0.21.0

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 REST API in one command, query it instantly, generate environment-aware automation scripts — no MCP servers, no code generation, no runtime complexity.

pip install olav
olav registry register http://netbox:8000        # connect any API
olav "how many devices are in NetBox?"            # query immediately
olav --agent netops "write a backup script"          # generate real scripts

Quick Start | Docs | 中文

---

Why OLAV?

API-as-Service — Beyond MCP

MCP requires a server process per service, stdio/HTTP transport, and framework-specific adapters. OLAV takes a different approach:

MCP:   Service → MCP server process → stdio/HTTP → adapter → agent
OLAV:  Service → olav registry register → reference markdown → api_request → done

One command. No server processes. No generated code. No runtime overhead.

# Register once
olav registry register http://netbox:8000

# Query from any agent, forever
olav "how many devices are in NetBox?"
olav --agent netops "compare OLAV database vs NetBox — are they in sync?"

The api_request tool is schema-aware — it reads API reference docs generated at registration time, handles pagination (DRF/NetBox style), and manages auth (JWT/Bearer/API-key) automatically.

Three Agents — Strict Tool Isolation

olav "list all devices"                  → Core Agent (database + knowledge base)
olav --agent netops "simulate link failure" → Ops Agent (network operations + SSH)
olav --agent audit "run health check"    → Audit Agent (compliance + learning)

Core agent uses a subagent architecture — the orchestrator sees only 5 tools:

core orchestrator (5 tools: execute_sql, olav_recall_memory, web_search, format_and_export, olav_delegate)
  ├── db_query    — database queries, knowledge base, web search, export
  ├── api_query   — API requests, health checks, web search, export
  ├── remote      — SSH to servers, local shell commands
  └── admin       — platform management, deployment, cron

Each subagent has only the tools it needs. execute_sql and api_request are in different subagents — the LLM cannot confuse them. Principle of least authority, enforced by the harness.

7-Layer Write Security

AI agents that can write to production need more than HITL approval:

Layer	Defense	Bypassable?
--enable-api-write	Write mode locked by default	No
services.yaml readonly_only	Per-service read/write control	Config only
Dry-run simulation	Must pass before approval offered	No
HITL approval	User sees diff, then confirms	Not skippable
sandbox_guard hard_block	HTTP writes in isolated sandbox	Not skippable
unshare --net	Kernel-level network isolation	Not skippable
Audit trail	Every api_request logged	—

--dangerously-skip-permissions bypasses tool approval for testing — but cannot bypass API write approval. Network devices are always read-only.

Agent Harness — The OS for AI Agents

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 interception + memory injection
Layer 2: Sandbox    Pre-scan → DuckDB read-only → network namespace isolation
Layer 3: Output     Credential redaction + SSE encoding

Self-Improving Loop

Use OLAV → audit log captures every tool call
    → failure patterns extracted → written to LanceDB memory
    → future runs recall constraints before acting

Export as SFT/trajectory training data: olav log export sft.

---

Quick Start

# 1. Install
pip install olav

# 2. Initialize
olav init

# 3. Configure LLM
nano .olav/config/api.json   # set shared.api_key + llm.model

# 4. Connect a service
olav registry register http://netbox:8000

# 5. Query
olav "how many devices are in NetBox?"

Network Operations (optional)

olav agent install /path/to/olav-netops/          # adds netops + devops workspaces
                                                   # (legacy alias: `olav skill install`)

olav --agent netops "/netops_init"                 # collect device data via SSH
olav --agent netops "simulate R2 link failure"     # What-If analysis
olav --agent netops "deploy digital twin"          # ContainerLab validation

Other Interfaces

olav                            # interactive TUI
olav service start --all        # web UI at localhost:2280
olav --agent core "run: df -h"  # shell commands via Core Agent

---

Architecture

olav v0.18.0 (pip install olav)
├── core orchestrator (5 tools)
│   ├── db_query    — execute_sql, olav_recall_memory, web_search, format_and_export
│   ├── api_query   — api_request, service_health, web_search, format_and_export
│   ├── remote      — remote_execute (SSH), run_shell
│   └── admin       — workspace_health, bulk_ingest, deploy/stop_service, cron, ...
│
olav-netops v0.19.0 (olav agent install olav-netops/)
├── netops orchestrator
│   ├── probe    — Parallel SSH with command whitelist (Nornir)
│   ├── analysis — Dijkstra + ECMP simulation (networkx)
│   ├── diff     — Cross-snapshot drift detection
│   └── lab      — ContainerLab digital twin + commit-validate
├── audit
│   ├── design   — Compliance profiles + health reports
│   └── learn    — TextFSM template learning
└── netops.*     — DuckDB tables + TextFSM collection pipeline

Tech Stack: LangChain · LangGraph · DeepAgents · DuckDB · LanceDB · FastAPI · NetworkX

---

Documentation

Docs: docs.olavai.com · Website: olavai.com

---

License

BSL-1.1 — Business Source License 1.1