Annoying and alarming things about OpenCode
Points and comments are a snapshot, not live.
OpenCode has severe security and usability flaws due to its design.
The post details numerous issues with OpenCode, a popular open-source AI coding agent. Annoying problems include prompt cache misses (e.g., from re-reading AGENTS.md each turn, including the current date), broken pruning that deletes specs from context, and a clumsy TUI. Alarming security issues are a remote-first design that connects to a remote model by default, a WebFetch tool for unpredictable LLM internet access, and a bash permission system that uses flawed AST parsing and regex matching.
What commenters are saying
Commenters largely agree with the critique, with some sharing experiences switching to alternatives like Pi. A key debate centers on alternatives for free inference, which the author cites as their reason for using OpenCode. Others note the project's 3,690 open issues and lack of PR acceptance, suggesting the project is unmaintained. One commenter found Pi offers better tool calling performance.