Open Source · Agent OS

Kronos Agent OS

A self-hosted Python runtime for long-running AI agents. Where agent frameworks optimize a single conversation, KAOS is the operational layer around the agent: persistent memory, skills, audited tools, scheduled automations, multi-agent coordination, and an operator dashboard — all included.

PythonReActMCPMulti-agentSQLiteSelf-hostedLLM
5
memory layers
5
security layers
MIT
open source

What an Agent OS is

The operational layer between the LLM and the user.

When an agent runs for weeks instead of minutes, the requirements change: inspectable state, tool-call auditing, scheduled tasks, security controls, multi-agent coordination. Most frameworks punt these to the developer. KAOS ships them as a runtime — self-hosted, no cloud service needed, managed through a local dashboard.

What's inside

The operating layer around a long-running agent.

Five-layer memory

Sessions, FTS5 full-text, vector search, a knowledge graph, and shared facts. The agent remembers across sessions — not just within one conversation.

Custom ReAct engine

~200 lines, no LangGraph. Full control over tool calling, timeouts, and the audit trail. Every tool call is logged with secret redaction.

Defense-in-depth security

Prompt-injection shield (28 patterns), input sanitizer, loop detector, output validator, and a cost guardian ($5/day default). Five capability gates, all off by default.

Swarm Mode

Six independent agents coordinate through SQLite IMMEDIATE transactions — deterministic arbitration without pub/sub. Optional: single-agent by default.

Scheduled automations

12+ built-in automations with progressive skill disclosure. The runtime keeps working on a schedule, not only when you prompt it.

Bring your own LLM

Any OpenAI-compatible endpoint — Fireworks, DeepSeek, OpenAI, OpenRouter, Groq, Together, Ollama. Set the URL and model in .env; no code changes.

Engineering quality

Kronos is a public, verifiable project — so its quality is measured, not asserted. Live snapshot from the canonical collector; the full methodology is on the engineering dashboard.

CI-generated
Python / public, verifiable

Kronos Agent OS

Production Python only; tests, virtual environments, migrations, and runtime caches excluded from headline metrics.

Tests
747
0 failed
Core coverage
53.96%
Format compliance
100%
Duplication
0.83%
Complexity
A
Static findings
0
Critical / high CVEs
0
pip-audit + Bandit
Type errors
281

Verified: 2026-07-16 · commit ea3ddc4

Run it yourself

MIT-licensed, pip-installable, with an offline demo that needs no API keys. Any OpenAI-compatible model.

View on GitHub

github.com/spyrae/kronos-agent-os