OpenCode

What was covered

• What it is. A command-line application with a terminal interface. It runs locally or over SSH on any machine, including servers and customer machines, and needs no administrator privileges, which helps in restricted corporate environments.
• Plan and build. Two main agents. Plan is for strategy and long-term changes, build is for implementation. The tip is to start in plan, review the plan carefully, then switch to build.
• Model freedom. Multiple providers (cloud platforms and the model vendors) and local models through Ollama. You pay for tokens instead of a subscription, and you can switch provider inside a session with a login command.
• Using an existing Copilot subscription. It can reach expensive models through a GitHub Copilot subscription, which lets you mix models and providers as you work.
• Project context. An agent.md file in the repository gives the agent the project’s context and conventions, so the whole team shares one setup. Custom agents can use different models and prompts.
• Skills and commands. Skills are shared instruction sets for a process, such as how to do a release update. Felipe prefers one or two agents with many skills over many agents. Markdown files such as test.md become custom commands with arguments.
• Language servers and sessions. LSP integration runs static analysers and formatters automatically on generated code. Each session is saved on its own, so work from months ago can be resumed.

What came up

• Versus Cursor. Felipe presented it as a stronger alternative and said he had stopped using and paying for Cursor.
• Intermediate layers. Asked whether a gateway between the model and the API could change behaviour, he said OpenCode understands such layers and routes requests correctly.
• Community. An open-source project from the Ollama ecosystem was suggested for exploring prebuilt agents and new skills.

Takeaway

The aim of the talk was to get people to try it rather than to list every feature.