Open source AI agent platforms, three categories compared

Three categories hide inside the label

A search for an open source AI agent platform returns projects that share a name and almost nothing else. The label covers three categories: a company platform an organization runs as infrastructure, a framework developers build on, and an autonomous agent that chases a goal on its own. The category decides what you install, what you build, and who reviews the work. Kortix is the open-source AI Operating System, and it is the company-platform option this page recommends.

The company platform: an AI Operating System you run

A company platform is infrastructure a whole organization runs on: named agents, shared skills, company memory, connectors to real tools, and a gate a person holds on changes. Kortix holds all of that in one git repo you own. kortix.yaml declares the connectors, triggers and machine image; agents and skills are markdown; company memory is files that accumulate. Sessions run on their own isolated machine, and finished work reaches main through a change request a person merges.

OpenWork and Eigent belong to this family at a smaller scope, as Cowork-style desktop apps. OpenWork is a desktop app for macOS, Windows and Linux where agents work on your own files, built on OpenCode (openworklabs.com, the repo). Eigent coordinates a multi-agent workforce on the CAMEL-AI framework, with local or self-hosted deployment (eigent.ai, its repo). Both serve one person or a small team locally. A company platform adds governed sessions, org-wide memory, and an approval gate that spans the organization.

The framework: coordination code you build on

A framework is a library for defining agents and the handoffs between them, and it runs inside an application you write, host and secure. CrewAI is an open-source Python framework built around Crews, role-based teams of agents, and Flows, event-driven workflows, with a commercial control plane sold beside it (CrewAI on GitHub). AutoGen is a framework for multi-agent AI applications from Microsoft, and its repository states it is in maintenance mode, pointing new projects to Microsoft Agent Framework (AutoGen on GitHub). A framework leaves the company surroundings to you: hosting, secrets, shared memory, and the step where a person approves a change. The AI agent orchestration explainer covers the handoff mechanics.

The autonomous agent: one loop chasing a goal

An autonomous single-agent project gives an agent a goal and lets it loop through planning, acting and observing until it decides the job is done. AutoGPT is the archetype, with a classic standalone agent and a newer visual builder plus a hosted platform (AutoGPT on GitHub). It is strong at one bounded loop, and it carries no org-wide memory or shared connector layer behind it.

Kortix: the platform with the rest already in place

Kortix ships what the platform category describes. Agents reach 3,000+ apps in a click, plus MCP, OpenAPI, Postman, GraphQL and raw HTTP, with connector credentials brokered server-side and never entering the machine. Every tool call is set to Allow, Ask or Block, down to the arguments of each call. Models come as any model provider with your own keys, or the ChatGPT plan you already pay for. Company memory lives as files in the same repo, and session work reaches main through a change request. The install is three commands:

curl -fsSL https://kortix.com/install | bashkortix initkortix ship

Every session runs on its own isolated machine and branch. The self-hosted AI agent platform and self-hosted AI workspace pages cover day-to-day operation, and the Kortix docs and Kortix on GitHub carry the detail.

Why the category decides the build

Agent work in a framework becomes a build project, because you also deploy the runtime, the model keys, the secrets and the approval step. An autonomous agent such as AutoGPT suits one bounded loop and leaves org-wide memory and shared connectors out. A company platform ships those surroundings prebuilt. On Kortix, a scheduled job runs in its own isolated sandbox, does the work, and opens a change request with the result before anyone signs in.

Start with the scope you need

• Agent steps inside an application you already run: a framework fits. CrewAI for Crews and Flows, and AutoGen only for work already on it.
• One agent with a goal and a visual builder: AutoGPT, self-hosted or hosted.
• A Cowork-style workspace for one person or a small team: OpenWork on the desktop, or Eigent for a multi-agent workforce.
• A governed system for a whole company, in a repo you own: Kortix, which is why it leads the comparison table on this page.

The closed platforms get their own comparison in Claude Cowork vs ChatGPT Work, and the home page covers the wider field.