Introduction

Kortecx is the agentic runtime. An agent is a program that uses a model to decide what to do next — read a document, call a tool, write an answer. Kortecx is the thing that actually runs those agents, keeps a durable record of every step, and lets you pick the work back up after a crash.

Start in the console: describe a task in plain language, and the runtime plans it, acts, and hands back a durable record of every step it took — which tool it reached for, what was refused, and what it answered.

Talk to us →

The seven things you get

• Apps you author — an App is the shareable unit: one agent, its workflow, its instructions, and its references to context, tools and data, packaged in a single file you can run, export and import.
• Workflows that run — a workflow is a set of steps with a shape. Steps can run one after another or side by side, and agents can hand work to other agents inside one App. Apps do not call other Apps.
• Tools you create — the actions an agent is allowed to take. Registering a tool records it; connections are the path that actually dials an external system and calls out.
• Context it remembers — reusable instruction and file bundles, datasets you can search by meaning (needs an hnsw build), and durable memory across runs (needs KX_SERVE_MEMORY=1, a served model, and an hnsw build).
• Work on a schedule — a run that starts on a clock or on an incoming event, instead of because you typed something.
• Served at an address — the runtime answers on a local address, so the console, the command line, and the Python and TypeScript SDKs all talk to the same thing over the same wire.
• Run models locally — a model can run on your own hardware instead of behind a remote API, so nothing leaves the machine. This is not on by default. You either install a local inference backend that the runtime detects on loopback, or build with the non-default in-process inference feature (Metal or CPU) and point the runtime at a model file you supply.

How it holds together

Every step an agent takes is written to an append-only log before its result is trusted. If the process dies, the runtime rebuilds its state by re-reading that log. A step that already touched the outside world is re-read from its committed result, never run a second time.

Two rules follow from that, and they shape everything else:

Kortecx is single-system by default — one machine, one log. Running across several machines is on the roadmap, not shipped.

Start here

If you do not write code, follow these six pages in order. Each one ends where the next one begins.

Prove it to yourself

If you write code, do not take the durability claim on faith. The quickstart runs a demo workflow, kills it mid-commit, replays it from the log, and shows you the same digest both times.

What to expect

Kortecx is pre-1.0. The durability spine and the agent loop work; surfaces are still being filled in. Pin a release tag for anything you keep.

For running it somewhere that matters, read production notes first. You get an audit trail and live event stream, an opt-in Prometheus /metrics endpoint, terminal failures landing in an inbox instead of vanishing, and a local spend estimate at rates you configure, with a ceiling — not a bill.