What a self-hosted AI workspace is, and the open-source way to run one

The two layers a self-hosted AI workspace hides

A self-hosted AI workspace is two systems behind one phrase, and most people mean only one of them. The model layer is the language model that reads a prompt and writes an answer, running as inference on hardware you control. The agent tooling layer is everything around that model: the harness that turns it into an agent, the memory the workspace accumulates, and the connectors that reach your tools.

Self-hosting the tooling and self-hosting the model are separate jobs with separate costs. The tooling is software you run yourself, meaning the platform, its database, its file storage and its app. The model is compute you run, or an API you call. A workspace can own one layer and rent the other, and many real setups do exactly that.

Kortix is the open-source AI Operating System: your agents, their skills, your company memory, and every connector in one git repo you own, with the agents working on real cloud computers. The harness is OpenCode, and it turns any model you connect into an agent that does the work.

Three arrangements, and the one Kortix is built for

Self-hosted AI workspaces fall into three arrangements. In the first, everything is local: a model runs on your own GPU through a local server, and the tooling runs on the same machine. In the second, the tooling is self-hosted and the model is hosted: you run the platform on your own box and call a provider’s API with your own key. In the third, the whole workspace is managed, and one vendor runs the model and the tooling on its cloud.

Kortix is built for the second arrangement. Run it on Kortix Cloud, in your VPC, or on your own on-prem network. Self-host is free. Any model provider with your own keys. The same wiring also reaches both extremes: an OpenAI-compatible endpoint in front of a local model server, or the managed Kortix Cloud.

What you can run today

Kortix keeps your agents, skills, memory, connectors and triggers in one git repo, and each session runs on its own isolated sandbox on its own branch. Session work reaches main through a change request. Merge is default-deny for agents, so a person reads the diff first. It reaches 3,000+ apps in a click, plus MCP, OpenAPI, Postman, GraphQL and raw HTTP, and connector credentials are brokered server-side and never enter the machine. The self-hosting guide walks through the install on your own box.

OpenWork is a free, open-source desktop app for macOS, Windows and Linux that runs agents on your own files. Its own page says it is built on OpenCode and can run self-hosted or as a managed private instance (openworklabs.com). In desktop mode, your files stay on your machine and prompts go to the model provider you connect.

Eigent is an open-source, general-purpose agent desktop app. Its self-hosting docs describe three choices: a managed application, a self-hosted setup with your own key, and a fully local mode on Ollama, vLLM, SGLang, LM Studio or LLaMA.cpp (Eigent self-hosting docs).

A managed workspace such as ChatGPT runs entirely on the vendor’s cloud with no self-host path, and OpenAI’s own page notes that chats may be reviewed and used to improve its models (openai.com/chatgpt).

The hardware reality

Self-hosting the tooling and self-hosting the model need very different hardware. Running Kortix is one Docker Compose stack on a Linux box, and the docs keep its default API memory limit of 640 MiB on an 8 GiB host (Kortix self-hosting docs). Agent sessions run on a separate sandbox provider rather than on that box, so the host does not grow with the number of sessions.

Running a model locally is the other budget. Eigent’s docs state that local model requirements depend on the model and runtime, and that large models can require substantial memory or GPU capacity (Eigent self-hosting docs). A tooling-only self-host on an 8 GiB box is small and quick. A fully local workspace needs a GPU with enough memory to hold the model weights.

Self-hosted workspace or a managed workspace

Managed and self-hosted workspaces trade control for setup. A managed workspace starts in minutes, runs on the vendor’s cloud, and keeps your prompts and files under the vendor’s terms. A self-hosted workspace asks for more setup, and in return you own the configuration, pick the model, and keep the memory in a repo you can read and back up.

Kortix leans toward ownership. It is the leading open-source alternative to Claude Cowork and ChatGPT Work. You can run the tooling on your own box, connect any model with your own keys, and still start on the managed cloud for the fast path. The self-hosted AI agent platform guide sets the two sides side by side, AI agent orchestration covers running many sessions at once, and Read the guide goes deeper on self-hosted workspaces.