Cloud Parts

OpenCode

An open-source (MIT) coding agent that works with 75+ model providers and local models, built as a local server with a terminal interface as one of its clients.

What it is
AI coding agents (terminal agents)
Type
Open source
Made by
Anomaly Innovations
License
MIT
Facts checked
9 October 2026

What it is

OpenCode runs in the terminal, as a desktop app or inside an editor. It isn’t tied to one model vendor: it uses the AI SDK and the Models.dev catalogue to reach more than 75 providers, plus models running locally. Two built-in agents switch with the Tab key: build, the default, with full access, and plan, which is read-only, denies file edits and asks before running shell commands. A general subagent handles searches and multi-step side tasks. It can start language servers for the files it touches and feed their diagnostics back to the model.

Under the interface is a client/server split. Starting opencode starts both a TUI and an HTTP server, and opencode serve runs the server alone, with an OpenAPI 3.1 spec at /doc that is also used to generate an SDK. IDE plugins drive the TUI through that server, and opencode acp speaks the Agent Client Protocol for editors that support it.

Instructions come from AGENTS.md (created by /init), found by walking up from the current directory, then a global ~/.config/opencode/AGENTS.md. For people moving from Claude Code it also reads CLAUDE.md and ~/.claude/skills unless told not to. Permissions resolve each action to allow, ask or deny, and start permissive: most actions are allowed, while touching paths outside the project and a tool call repeated three times with identical input (doom_loop) ask first, and .env files can’t be read.

Why teams pick it

  • Provider-agnostic: swap between hosted models, existing subscriptions and local models without changing tools
  • MIT-licensed, with a documented HTTP API, so you can script it or build your own client
  • Reads CLAUDE.md and Claude Code skills, which makes switching cheap
  • Language-server diagnostics give the model compiler-grade feedback after each edit

Watch out for

  • Permissive defaults: most actions run without asking unless you add ask or deny rules
  • Results depend on the model you pick; there is no single vendor tuning model and harness together
  • The server is unsecured unless you set OPENCODE_SERVER_PASSWORD, so be careful binding it beyond 127.0.0.1
  • Easy to confuse with the archived Go project opencode-ai/opencode, which now continues as Crush under Charm

The OpenCode family

1 technologies built from or on OpenCode

Alternatives

What all terminal agents have in common →

Run it in a terminal, one-shot, or as a server · bash

curl -fsSL https://opencode.ai/install | bash   # or: npm i -g opencode-ai

opencode                         # TUI; Tab switches between the build and plan agents
opencode run "explain how the auth middleware works"   # one message, then exit

# headless server: the same API the TUI and IDE plugins use
opencode serve --port 4096 &
curl -s http://127.0.0.1:4096/global/health
# {"healthy":true,"version":"1.18.35"}

See it in a real stack

No architecture uses OpenCode yet. Drop it into the playground and wire it up yourself.Open playground →
Where it came from2025–2026

OpenCode’s repository started in spring 2025 under the GitHub organisation of SST, and published its first releases in May 2025. Until June its README described a Go program with a Bubble Tea terminal interface and promised “a complete overhaul” for mid-June; the rewrite that followed is written in TypeScript. Another Go project called OpenCode, opencode-ai/opencode, has been archived and continues as Crush, developed by its original author and the Charm team; OpenCode’s own FAQ says that repository has no relation to it.

Its FAQ set out the pitch against Claude Code: similar capability, but fully open source, not coupled to any provider (Anthropic was recommended, but OpenAI, Google and local models work too), language-server support out of the box, a terminal interface built by Neovim users and the creators of terminal.shop, and a client/server design so the TUI is just one possible client. Version 1.0.0 shipped on 31 October 2025. By February 2026 the repository had over 95,000 stars, and it now lives under Anomaly’s GitHub organisation.

Why it’s built this way

Not coupled to one provider
Its authors expected the gaps between models to close and prices to drop, so being able to switch providers mattered more than tuning for one.
A server first, the TUI as a client
Separating the agent from its interface lets it run on one machine while being driven from another, such as a phone, and lets IDE plugins and scripts use the same API.
Language servers in the loop
Diagnostics from a language server tell the model about type errors right after an edit, without waiting for a build or test run.
Compatible with Claude Code’s files
Reading CLAUDE.md and Claude Code skills by default lowers the cost of trying it in a repository already set up for another agent.

Timeline

  1. 2025

    First releases under SST’s GitHub organisation (May)

  2. 2025

    The Go version is replaced by a rewrite announced for mid-June

  3. 2025

    Version 1.0.0 (31 October)

  4. 2026

    Passes 95,000 GitHub stars; published by Anomaly Innovations

Sources9 · facts checked 9 October 2026

Improve this page on GitHub