Start with instructions. Add code when you need it.
Describe a task, choose the actions your agent can use, and try it locally. You can start with a single Markdown file.
Your first agent can be one file.
For example, create an agent that helps triage GitHub issues. Reuse the GitHub plugin’s actions so you can focus on the instructions.
oz0 agent new triage --uses github1. Describe the task
Edit the generated Markdown file with your instructions and the tool actions the agent may use.
2. Try it locally
Run
oz0 dev. Your agent reloads when you save.3. Grow when needed
oz0 plugin ejectturns it into a full folder for code, tests and more parts.
oz0’s default runtime manages context and the model → tool → result loop. You choose a compatible model and permitted tools. An external harness is optional; control coverage varies by runtime.
Add only what the task needs.
- Need a new action?
- Add a tool: code the agent can call to do something.
- Need a repeatable procedure?
- Add a skill with the instructions for that procedure.
- Need to check the result?
- Add evals: test cases that check whether the agent did the job well.
Build with the tools you already use.
oz0 setup gives your coding assistant instructions for building and checking plugins. The kit works with Claude Code, Codex, Cursor and other coding assistants.
- Write new code
- Use Python or TypeScript. Other languages can run as a remote tool server or container. Tool options
- Reuse an API or tool server
- Import an OpenAPI description or connect an existing MCP server. Import guide
- Bring an existing agent
- Adapt a script, LangGraph graph or CrewAI crew. Agent integration guide
- Choose a model
- Connect a compatible hosted or self-hosted model. Reuse your agent and tools when switching; evaluate the new model on your tasks. Model options