// the find
anomalyco/opencode
The open source coding agent.
opencode is a terminal-based AI coding agent — a direct alternative to Claude Code, built in TypeScript with a build/plan agent split and a subagent model for complex tasks. It ships as a CLI, a beta desktop app, and a VS Code/GitHub Action integration, aimed at developers who want an open-source agent they can run against any model provider rather than being locked into one vendor's tool.
Distribution is taken seriously: install script plus npm, brew, scoop, choco, pacman, nix flakes, and mise, with nix/hashes.json and normalize-bun-binaries.ts suggesting real effort toward reproducible builds rather than just a curl-bash script. The build/plan agent split is a sane default — plan is read-only and asks before running bash, which is the right failure mode for exploring an unfamiliar codebase. The e2e suite goes further than most projects in this space: there's a dedicated visual-stability testing framework (analyzer, invariant checks, regions) built specifically to catch UI flicker/regressions in the timeline view, which is not something you see in typical TUI projects. The repo dogfoods itself hard — .opencode/agent and .opencode/command contain real automation (triage, duplicate-PR detection, changelog generation) that presumably runs opencode against its own repo.
The README is pure install-and-marketing copy — zero explanation of how the agent loop, tool permissions, or provider abstraction actually work, which is a strange gap for a project whose whole pitch is being the open alternative worth understanding. The artifacts/ directory has committed .mp4 and .png files (demo videos, generated clips) sitting in the source tree, which bloats clone size for no functional reason. 30+ GitHub Actions workflows plus SST-based infra (infra/app.ts, infra/console.ts, infra/enterprise.ts) signal a genuinely complex deployment surface — self-hosting or forking this isn't a weekend project, and none of that operational complexity is documented in the README. The packages/app directory alone has a sprawling e2e/regression suite with 30+ spec files and no visible top-level architecture doc pointing you to where core agent logic actually lives.