Player not loading? Watch on YouTube
This comparison focuses on how a coding assistant handles work around the model: context, permissions, edits and tool use. The speaker recommends OpenCode as a starting point, Aider for reviewed Git changes, and Pi for developers who want to build their own agent loop.
OpenCode is described as MIT-licensed, with local model connections through Ollama or an OpenAI-compatible endpoint. Its client-server architecture supports headless use. The speaker says Claude access requires an API key because Anthropic prohibits using Pro and Max subscriptions through OpenCode.
Aider's workflow centers on atomic commits and a repository map. It can switch models during a session and use local models for offline work. The stated limitation is no native MCP support as of mid-2026; the speaker presents it as a pair programmer that needs ongoing review.
Pi uses a small system prompt and loads skills when needed. The speaker describes it as less polished than OpenCode but easier to customize. The hardware section discusses four-bit quantized models on a 24 GB GPU and vLLM as a serving layer for parallel agents. Claude Code provides a proprietary comparison, with hooks and sub-agent workflows, but the speaker says it is neither local nor open.
