What Is OpenCode?

OpenCode is an open-source coding agent. You point it at a project and a model, and it reads your code, edits files, runs commands, and works through a task with you. It runs as a terminal UI, a desktop app, a web UI, and an extension for VS Code, Cursor, Zed, and similar editors. It's built by Anomaly and released under the MIT license.

Is it free? The agent is. OpenCode's FAQ says it's "100% free to use." Model usage is a separate question: you either bring a provider you already pay for, use OpenCode's limited free models, or pay OpenCode for model access through Go or Zen/Console.

Which models? Almost any. OpenCode connects to 75+ providers through Models.dev, including Anthropic, OpenAI, Google Vertex, Bedrock, OpenRouter, and local models through Ollama, LM Studio, and others. You can also sign in with a ChatGPT Plus/Pro or GitHub Copilot subscription.

OpenCode isn't OpenCode Go. Go is a separate $10/month model subscription you can plug into OpenCode. Zen, which OpenCode's newer docs call Console, is its pay-as-you-go model gateway. Neither is required.

Which version? As of September 26, 2026, OpenCode's homepage and download page install v2, while GitHub Releases and the changelog still ship v1 (1.18.32 on September 21). Most third-party install guides give the v1 commands. Details below.

Which repo? The current code is at github.com/anomalyco/opencode. The opencode-ai/opencode repo that tops several GitHub searches is a different, archived project that became Charm's Crush.

OpenCode vs OpenCode Go vs Zen/Console

These share a name, an account, and a website, so search results and reviews mix them up. They're separate products.

On the naming: OpenCode's Zen product page and v1 docs say "Zen." The v2 docs call the same service "OpenCode Console" and describe it more broadly, as an optional account service for teams that covers pay-as-you-go inference, bring-your-own-provider routing, budgets, team policies, web search, and Go. The API endpoints still sit under opencode.ai/zen/, and model IDs still use opencode/<model-id>. We treat Zen and Console as one service mid-rename. OpenCode hasn't published a rename announcement that we could find.

OpenCode Black also has a page at opencode.ai/black ("Access all the world's best coding models"), but it shows no price and says "Black plan enrollment is temporarily paused." We don't cover it further.

Go has enough moving parts (per-model limits, privacy that varies by model, a roster that changes often) that it has its own page.

OpenCode Go: what the $10 plan includes, its current model roster, how the limits work, and privacy by model

product↗

Is OpenCode free?

The software is free and open source. Whether you pay anything depends on where your model tokens come from.

Free models aren't free of tradeoffs. OpenCode's FAQ says your data "is only stored when you use our free models or create sharable links." The Zen docs say that for Big Pickle, "collected data may be used to improve the model" during its free period, and some other free models carry their own data terms. Don't send code you can't share to a free model without reading its entry in the Zen docs.

Install OpenCode

OpenCode's homepage and download page currently install v2:

curl -fsSL https://opencode.ai/v2/install | bash

The v2 docs also list npm install -g @opencode/cli, brew install anomalyco/tap/opencode-v2, and paru -S opencode-beta on Arch. There are standalone binaries for macOS, Linux (glibc and musl), and Windows (x64, ARM64, and a baseline build). The v2 docs say "Windows package managers are not supported."

Then open your project and start it:

cd your-project
opencode

In the TUI, /connect adds a provider (an API key, a subscription sign-in, or OpenCode's own models), and /models picks a model. Instructions for the agent go in AGENTS.md at the project root.

The v1 commands still work, and still install v1. The GitHub README, the v1 docs, and most third-party guides give curl -fsSL https://opencode.ai/install | bash or npm install -g opencode-ai. On September 26, that script redirected to the v1 installer on GitHub, which downloads the latest v1 release. Both installers write the same opencode binary path, so running one replaces the other. OpenCode's migration guide says to remove a package-managed v1 install before installing v2.

Desktop and editors. The desktop app has builds for macOS (Apple silicon and Intel), Windows, and Linux (.deb, .rpm, and AppImage), plus brew install --cask opencode-desktop. The GitHub README labels it "BETA"; the download page doesn't. The IDE extension covers VS Code, Cursor, Zed, Windsurf, and VSCodium, and OpenCode's IDE docs say it installs itself when you run opencode in the editor's terminal.

Windows. The v1 docs recommend WSL "for the best experience." The v2 docs link native Windows binaries and don't mention WSL.

Phones. OpenCode's download page doesn't list a mobile app. The "OpenCode: AI Coding Agent" Android app that ranks for "opencode" is a third-party client for a self-hosted OpenCode server, published by Vibe Technologies, LLC.

OpenCode v2 vs v1: what to follow today

OpenCode is shipping two lines at once, and its own pages point in both directions.

OpenCode's v2 migration guide lists three intentional breaking changes: plugins use a new API ("V1 plugin implementations do not run in V2"), the server API and clients have new contracts, and terminal settings move from tui.json files to one global cli.json, which is migrated automatically. It says existing config, agents, commands, and skills "should continue to work without changes," and that v2 reads v1-format config in place.

The same guide, plus the v2 config docs, list gaps you should know about before switching:

• No language servers. v2 "accepts and preserves lsp configuration, but it does not run language servers."
• No CLAUDE.md fallback. v1 reads CLAUDE.md when there's no AGENTS.md. "V2 currently only discovers AGENTS.md."
• No session sharing yet. The v2 config docs say sharing settings are accepted "but session sharing is not supported yet."
• No Windows package managers, and no built-in scout subagent.

Models and providers

OpenCode doesn't come with one vendor's model. You connect providers, and it uses the AI SDK and Models.dev to support "75+ LLM providers." Credentials you add with /connect or opencode auth login are stored locally, and OpenCode also picks up provider keys from your environment.

The providers docs have dedicated sections for Anthropic, OpenAI, Google Vertex AI, Amazon Bedrock, Azure OpenAI, GitHub Copilot, GitLab Duo, OpenRouter, xAI, DeepSeek, Groq, Together AI, Vercel AI Gateway, and dozens more. Anything else with an OpenAI-compatible API can be added as a custom provider with a baseURL. You choose a model per session with /models, per agent in config, or with --model, written as provider/model, like anthropic/claude-sonnet-4-5.

OpenCode's own model list is short on advice. Its docs say only "a few" models "are good at both generating code and tool calling," and the recommended list on the v1 models page still names models that its Zen docs mark as deprecated. For choosing a model, OpenCode's docs aren't the place to decide.

How to choose a model for a task by reasoning depth, latency, and cost before wiring it into an agent

concept↗

Local models

OpenCode supports local models, but it doesn't make them easy or equal to hosted ones.

Those two numbers conflict, and neither source says which models work well. Agents send long prompts and depend on tool calling, so small local models with default context windows often fall short. That's an inference from how agents work, not something we've measured with OpenCode.

Agents, MCP, skills, plugins, and permissions

OpenCode is closer to a configurable agent platform than a single chat loop. Here's what ships, per the v1 and v2 docs:

• Agents. Two primary agents you switch between with Tab: Build, the default, with tools allowed, and Plan, which explores and plans without editing project files. Subagents include General and Explore (v1 also has Scout for external docs). You can define your own agents as Markdown files in .opencode/agents/.
• MCP. Local servers over stdio and remote servers over HTTP, with OAuth handled automatically for remote ones. The docs warn that MCP servers "add to the context," and that some, like GitHub's, "can easily exceed the context limit."
• Skills. SKILL.md folders in .opencode/skills/. OpenCode also reads .claude/skills/ and .agents/skills/, so existing Claude Code skills load without copying.
• Commands. Custom slash commands as Markdown files in .opencode/commands/.
• Plugins. JavaScript or TypeScript modules, from local files or npm, that hook into events and can add tools. v1 plugins don't run in v2.
• Other surfaces. Headless runs for scripts, an HTTP server and SDK, Agent Client Protocol support for editors, and a GitHub Actions integration you trigger by commenting /opencode or /oc on an issue or PR.

Permissions are permissive by default. The v1 docs say "Most permissions default to allow," with ask for directories outside the project and for repeated identical tool calls, and .env reads denied. v2's base policy is also allow, with ask for outside directories and .env files. If you want approval before edits or shell commands, you configure it, or work in Plan.

The official GitHub repository

That archived repo still came first for "opencode github," "opencode cli," and "opencode ai" in the live search results we checked on September 26. Its README describes a Go and Bubble Tea app with a different feature set, and some third-party explainers repeat its description. If a page describes OpenCode that way, it's describing the old project.

One more source to be careful with: open-code.ai, which ranks for several OpenCode queries, calls itself "a community-maintained documentation site, not official." OpenCode's own docs are at opencode.ai/docs (v1) and opencode.ai/v2/docs.

How OpenCode differs from Claude Code, Codex, and Cursor

The factual differences are structural, not about quality:

• Model choice. OpenCode works with any supported provider. Claude Code is Anthropic's agent and Codex is OpenAI's, each built around its maker's models. With OpenCode you pick, and pay, the provider.
• License. OpenCode's code is MIT-licensed, and you can read, fork, and change it.
• Form factor. Cursor is an AI code editor. OpenCode is an agent that runs in a terminal, a desktop app, a browser, or inside an editor through its extension.
• Existing config. OpenCode reads .claude/skills/, and v1 falls back to CLAUDE.md, so some Claude Code setup carries over.

We haven't run OpenCode against any of them, so we're not saying which one does better work. Model choice changes results more than the harness name does, and the same model can behave differently in two agents. We'll publish a comparison when we have our own test results.

Where it fits in a coding-agent workflow

OpenCode covers the same loop as other coding agents: inspect the repo, plan, make bounded edits, run the checks. Plan mode is the built-in approval gate. The permissive defaults mean you should add your own gates, like ask on shell commands or a Plan pass before Build, when you point it at code you care about.

The production coding-agent workflow: scope, plan approval, bounded edits, deterministic checks, and a separate factual QA pass

workflow↗

What we haven't verified

• Whether the v2 installer, desktop app, or IDE extension work as documented on any platform. We didn't run them.
• When v2.0.0 was released and whether OpenCode considers v2 stable. We found no announcement or status label.
• Whether Claude Pro/Max sign-in works in any current build. OpenCode's FAQ and docs disagree.
• How well local models work in OpenCode, and what context size they actually need.
• How OpenCode compares with Claude Code, Codex, Cursor, or any other agent on real tasks.

How we separate vendor documentation, community reports, and our own inference

concept↗

Change log

Sources & Documentation

1. OpenCode: home page and FAQ — Product description, v2 install command, 75+ providers, ChatGPT and Copilot sign-in, "100% free to use," free models, Claude Pro/Max FAQ answer, privacy statement.
2. OpenCode: Download — v2 terminal install commands, desktop builds, IDE extensions, GitHub and GitLab integrations. No mobile app.
3. OpenCode v2 docs: Intro — v2 install commands, standalone binaries, desktop and web app, "Windows package managers are not supported."
4. OpenCode v2 docs: Migrate from V1 — Three breaking changes, shared binary and config locations, LSP and CLAUDE.md gaps.
5. OpenCode v2 docs: Config — Config locations and format; sharing "not supported yet."
6. OpenCode v2 docs: Models — Automatic discovery of Ollama, LM Studio, and vLLM models.
7. OpenCode v2 docs: Agents, MCP servers, Skills, Plugins, Permissions — Default agents, MCP transports, skill discovery including .claude/skills, v2 plugin API, base permission policy.
8. OpenCode v2 docs: Console — Console as the optional account service, including pay-as-you-go inference and Go.
9. OpenCode v2 update endpoint — Returned v2 CLI 2.0.18 (@opencode/cli) on September 26, 2026.
10. OpenCode docs (v1): Intro — v1 install commands, WSL recommendation, /connect and AGENTS.md, v2 banner.
11. OpenCode docs (v1): Providers — 75+ providers, credentials, Anthropic subscription warning, Copilot and GitLab Duo, Ollama, LM Studio, llama.cpp, custom providers.
12. OpenCode docs (v1): Models — Model selection and the outdated recommended-models list.
13. OpenCode docs (v1): Agents, MCP servers, Skills, Plugins, Permissions, Rules, Share, GitHub, IDE — Build and Plan agents, subagents, MCP context warning, permission defaults, CLAUDE.md fallback, share behavior, GitHub Actions share default.
14. OpenCode docs: Zen — Pay-as-you-go gateway, free models and their data terms, "completely optional."
15. OpenCode: Go product page and FAQ — Free-model quota of 200 requests/day; Go pricing.
16. OpenCode: Black — "Black plan enrollment is temporarily paused." No price shown.
17. OpenCode: Changelog — Only 1.x releases; v1.18.32 on September 21, 2026.
18. GitHub: anomalyco/opencode — Current repository, MIT license, README install commands, desktop app beta, CONTRIBUTING rules, v1.18.32 as the latest release.
19. GitHub: opencode-ai/opencode (archived) — Archived September 18, 2025; continued as Crush by Charm.
20. Ollama docs: OpenCode integration — Says OpenCode needs a 64k+ context window; uses v1 install commands.
21. Google Play: OpenCode: AI Coding Agent — Third-party Android client by Vibe Technologies, LLC.
22. open-code.ai (unofficial) — Cited only because it ranks for OpenCode queries; it describes itself as community-maintained and not official.