opencode
Top pickThe MIT terminal agent that picks up your CLAUDE.md where you left it.
opencode (MIT, maintained by Anomaly) is the open-source agent closest in shape to Claude Code: a terminal UI with a full-access `build` agent, a read-only `plan` agent and a `general` subagent, plus a desktop app in beta. It is provider-agnostic, with 75+ providers through the AI SDK and Models.dev, including Anthropic on an API key and local models via Ollama, LM Studio, llama.cpp or any OpenAI-compatible endpoint. For anyone coming from Claude Code it reads CLAUDE.md, ~/.claude/CLAUDE.md and ~/.claude/skills as fallbacks when no AGENTS.md exists, so an existing repository works on day one. Two optional paid services sit beside the free tool: OpenCode Go, a $10/month subscription to open-source models (Go Plus $40/month), and Zen, a pay-as-you-go gateway of tested models. Your own keys cost nothing extra.
What it does well
- +Reads CLAUDE.md, ~/.claude/CLAUDE.md and ~/.claude/skills as fallbacks, the lowest-effort move from Claude Code
- +75+ providers: keep Claude on an API key, switch models per task, or run fully local
- +The same plan-then-build rhythm: a read-only plan agent that denies edits by default and asks before running bash
- +The largest community in the category, about 211k GitHub stars, with commits daily
Where it falls short
- −Keeping Claude models means per-token API billing: opencode's own docs say Anthropic prohibits using Claude Pro/Max plans through it, and it stopped bundling the plugins that did so in version 1.3.0
- −The desktop app is still in beta; the terminal is the mature surface
- −A large configuration surface (providers, models, agents, permissions) is more to learn than one vendor's defaults
- −A local model is a real hardware ask and will trail a frontier model on hard multi-file work
Hardware note
The client is a single binary that runs on modest hardware, including over SSH. Running the model locally is the demanding part: Ollama's qwen3-coder:30b is a 19 GB download with a 256K context window, so roughly 24 GB of VRAM or a 32 GB+ unified-memory machine is a sensible floor, and Ollama's coding-agent guide recommends a context length of 64k or more for larger repositories.
opencode as an alternative to
Where opencode shows up in our comparisons, and how it ranked.
opencode head-to-head
Straight comparisons against the tools people weigh it against.