Skip to content
Open Source AI Agent Framework
SpecificationOpen Source AI Agent FrameworkRev 2026-10-14

How to build an open-source AI agent that hands back finished work

To build an open-source AI agent, you do not need to write the runtime yourself. Kortix is the open-source AI Operating System: install the CLI, scaffold a project, describe the agent as a markdown file, and start a session. The agent works on its own cloud computer and opens a change request you review.

From install to a merged change request

Five commands from the Kortix quickstart, in the order you run them.

Exhibitterminal
$ curl -fsSL https://kortix.com/install | bash$ kortix login
Exhibitterminal
$ kortix init my-agent$ cd my-agent
Source: https://kortix.com/docs/quickstart
Exhibitterminal
$ kortix ship
Source: https://kortix.com/docs/quickstart
Exhibitterminal
$ kortix sessions new --prompt "Add a skill that drafts the weekly status report"
Source: https://kortix.com/docs/quickstart
Exhibitterminal
$ kortix cr ls$ kortix cr diff 1$ kortix cr merge 1
Source: https://kortix.com/docs/quickstart

What building an agent means

An open source AI agent framework usually gives you a library and the pieces to wire together, and the application that runs them is yours to write. On Kortix, you describe the agent as a file and let the platform run it. In Kortix, an agent is a markdown file at agents/<name>.md that holds its prompt and a scoped reach into tools, plus an agents.<name> block in kortix.yaml that carries its governance. The platform supplies the harness, the sandbox and the review path, so your work is the agent's instructions and its reach.

From an empty directory to a running agent

The quickstart runs five steps. Install the CLI, and kortix init my-agent scaffolds a project with kortix.yaml plus starter agents, skills and runtime config. kortix ship validates the config, commits, pushes and brings the project live in the cloud. kortix sessions new starts a session, one unit of agent work on its own cloud computer and its own branch. The code panels on this page carry the exact commands for each step, from install through the change request.

What each agent needs

Four things turn a prompt into a working agent: connectors, memory, triggers, and permissions. Connectors reach 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. Company memory is plain files today that accumulate as the agent works. A trigger is a cron schedule or a signed webhook that starts a session with no person present. Access is per resource for people and agents, with roles, groups, and an audit trail. Every action is recorded. Each connector action is set to Allow, Ask or Block. An Ask holds the call until a person approves it, then the agent resumes.

Where the change lands and who reviews it

Session work reaches main through a change request, and merge is default-deny for agents. The agent commits, pushes its session branch, and opens the change request against the project's default branch; a person with the project.gitops.merge permission reads the diff and merges it. Each session runs on its own isolated machine and branch, so thousands can run in parallel without touching each other. The patterns that make that handoff reliable are covered in building reliable AI agent workflows, and the runtime a session uses is covered in running an agent framework.

How to self-host

Kortix runs on Kortix Cloud, in your VPC, or on your own on-prem network, and self-host is free. After installing the CLI, kortix self-host start brings up the full stack on your own box as one Docker Compose stack. The Kortix docs and Kortix on GitHub cover the rest.

Building an agent on Kortix

Do I need to write code to build an agent?

Not much. An agent is a markdown file with instructions plus a governance block in kortix.yaml, so most of the work is writing the prompt and choosing tools. You can start a session from the web app with nothing installed. Code matters when you write a skill or a custom connector.

Which models can it use?

Any model provider with your own keys, or the ChatGPT plan you already pay for, or sign in with the OpenCode Console account for OpenCode Zen and Go. Pick the model per agent, per session or per message. Kortix is model-agnostic, so you can switch the day a different one lands.

Can I self-host it?

Yes. Run it on Kortix Cloud, in your VPC, or on your own on-prem network. Self-host is free. Install the CLI and run kortix self-host start to bring up the full stack on your own box.

Where does the agent's work land?

The agent commits its work to its session branch and opens a change request against the project's default branch. A person reads the diff and merges it. Merge is default-deny for agents, so nothing reaches main without review.

Build your first agent on the open-source AI Operating System

Start in the web app with nothing installed, or take the CLI path from install to change request.