a minimal, hackable agentic framework for autonomous AI agents

Runs locally with Ollama, in the cloud with OpenRouter and ZAI. Extensible via a manifest-based plugin system. Inspired by the architecture of OpenClaw, rebuilt from scratch for local-first operation.

pkg dev dl
$ pip install agentkthx

Boot Sequence

From a fresh shell to a running agent in under a minute. Pick a backend — local via Ollama, or cloud via OpenRouter.

agentkthx · /dev/tty1

Or try it instantly in the browser via Colab — open the AgentKthx.ipynb notebook ↗

Backends

Hot-swap model providers via a single CLI flag. Local-first by design, cloud when you need it.

🏠
Ollama
local
OpenRouter
cloud
ZAI
cloud
🦙
llama-server
local
📦
BitNet
local

Capabilities

Multi-backend

Run agents locally with Ollama, in the cloud with OpenRouter and ZAI. Hot-swap models via a single CLI flag — no code changes.

ollama openrouter zai llama-server bitnet

Plugin system

Manifest-based, drop-in plugins for backends and features. v0.2 spec adds sandboxed tooling and lifecycle hooks.

manifest.json v0.2 spec turboquant acp

Souls

Identity + style + behavior bundles. Ship a soul for your use-case — helper, skills, or trading.

nova-helper nova-skills nova-trading

Local-first

Built around local inference. No telemetry, no phone-home. Your prompts, your model, your machine.

no-telemetry offline-capable PEP-668 aware

Skills engine

Built-in skill loader. Compose agents from reusable markdown-driven skills, or write your own with the bundled skill-creator.

crypto-signals codebase-audit test-harness skill-creator

Resilient

API resilience layer, persistent memory, tool cache, error recovery. Built to survive long-running loops.

retry/backoff tool-cache persistent-memory

README · live from GitHub

The README below is fetched from the GitHub repo and rendered server-side into a cache.

Documentation

Full docs hub — architecture, plugin spec (v0.2), API references, changelog. All server-cached, all live from the repo.