agentos ● live

agentos is the plumbing the rest of my work stands on. It's a small, terminal-native runtime that gives an agent the essentials: a bus of tools it can discover and call, pluggable LLM providers, persistent memory, and a tight event-driven loop.

It has since grown into a full agent platform I run from my phone. One OpenAI-compatible provider puts an entire model catalog — 120+ models — behind a searchable picker, and a ReAct loop lets the agent actually do things: search the web, read and write files, run Python or a shell, analyze images, and roughly twenty other tools. Two modes: a plain streaming chat with the human firmly in the loop, and an agent mode that streams its tool-by-tool reasoning live, with a stop button and a configurable step budget.

Because an agent with real machine access is dangerous, the safety model is the point rather than an afterthought. A permission layer runs every risky action past me — approve or deny, right on the phone — and a filesystem scope fences the agent into a chosen directory. Secrets never touch the source; a hashed, timing-safe PIN guards the door; it runs privately on my own network, not the open internet.

The part I'm fondest of: the agent can extend itself. When no tool fits, it writes a new one as a small JavaScript function that runs in a hardened sandbox and joins its toolset for the session. It also carries a library of ~88 capability guides it can search and follow, and its system prompt is regenerated every turn with live context — the model it's running as, the time of day, its scope, how many tools and skills it holds — so it is genuinely aware of its own environment. This is engineering, not a research claim: the runtime tool-forging sandbox and the skill catalog are adapted from the open-source @framers/agentos project (Apache-2.0); the kernel, the platform, and the safety layer are my own.