Axon
The platform to build agents. A terminal, a runtime, and a toolkit for building, running, and deploying reliable AI agents.
📖 Documentation · 🔍 Registry · 🌐 axon.arclabs.it
curl -fsSL https://axon.arclabs.it/install | bash
Installs Bun (if it isn't already present) and Axon. Then run axon — on first launch
you get Zeno, a real agent cloned from the registry, because an agent is the only
thing that can run and a new user has none.
What Axon is
Most agent frameworks hand you a loop to wire up. Axon doesn't. You author four things — identity, tools, scripts, and policy — and the runtime owns everything else: context assembly, tool dispatch, stop conditions, session persistence, and enforcement.
An agent is a folder. A standard Bun TypeScript project with a shape Axon knows how to read:
my-agent/
├── axon.config.ts what it loads, what it's allowed to do
├── src/
│ ├── boot.vue who the agent is — the standing prompt
│ ├── tools/ export a function, the agent can call it
│ ├── prompts/ reusable context, loaded per task
│ └── scripts/ orchestration — the primary authoring unit
└── server/ routes and hooks, so the world can reach it
No registration, no schema definitions, no wiring. Export a function from src/tools/
and it becomes callable.
Why it's different
Agents run in a box. An Axon agent is its own OS process with its own namespaces and filesystem view. The kernel mediates every machine-facing operation and enforces policy you declare — a profile sets your machine's ceiling, an agent narrows within it. The provider credential never enters the box.
Models are a property of the agent. A cognet names engines by role, never by model. Which model fills a role is resolved at boot from the providers you declared, so swapping providers changes how an agent thinks, never how it works. Anthropic, OpenAI, Google, Groq, Cerebras, DeepSeek, Mistral, Moonshot, OpenRouter, Perplexity, xAI, Z.AI, and local models via Ollama.
Everything is recorded. One durable append-only session log per run. The terminal, the Fleet VS Code debugger, and an attached cloud deployment all render the same run from the same events — including the exact document the model received.
Deploy is one command. axon deploy gives the same folder that ran in your terminal a
stable production URL, an API key, and durable storage. No Dockerfile, no infra config.
Quick start
axon # open the terminal
axon init my-agent # scaffold a new agent
axon dev # run locally with hot reload
axon <agent> -p "do a thing" # one instruction, then exit
axon deploy # ship it to Axon Cloud
Extend it
Install capabilities from the registry — additive, and nothing you wrote changes:
axon install @axon/discord # your agent becomes a Discord bot
axon search obsidian # find modules, agents, cognets, benches
Documentation
Everything is at axon.arclabs.it/docs/v2.
| Installation | Get running in about a minute |
| Your First Agent | The runtime loop, from the inside |
| The Agent Folder | What you author, and where |
| Kernel & Policy | How confinement actually works |
| Providers | Bring your own keys, or use managed inference |
| Deploy | One command to production |
| CLI reference | Every command |
Status
Early beta. The core runtime and the terminal are stable; capabilities are still expanding. Feedback is genuinely wanted — open an issue or find us via axon.arclabs.it.
License
Proprietary. © ArcLabs