OpenCode review

OpenCode is an open-source coding agent that runs on your machine, not a hosted coding API or a hosted-only SaaS app. It sits alongside your codebase: use its desktop app, terminal interface, mobile app or SDK to direct agents working on a project. The homepage lists desktop builds for macOS, Windows and Linux, a terminal UI, an iPhone app through TestFlight and Android as coming soon.

The agent connects to more than 200 LLM providers through Models.dev, including local models. You can start with free models without an account, bring your own subscription, or use OpenCode Go. This lets OpenCode serve as the interface and agent layer while model access comes from a provider or local model. It supports parallel sessions on the same project, MCP servers for additional tools, plugins, custom agents and permission controls. That is a broad integration surface, but it is not a server product that you deploy as a separate hosted service.

OpenCode offers an SDK for driving the agent from scripts, CI jobs and other tools. Its homepage links to SDK documentation but does not name supported programming languages or describe an API protocol. Before wiring it into a build system, confirm supported languages, methods, versioning and error handling in the vendor’s SDK docs.

For interactive use, the terminal UI works in any shell, editor or remote server. The homepage lists installation through a curl script, npm package @opencode/cli, Bun, Homebrew and paru. Desktop downloads cover macOS, Windows and Linux. The integration surfaces also include plugins that hook into events, local and remote MCP servers, reusable agent skills in SKILL.md files, and custom slash commands. OpenCode lists GitHub, Linear, Figma, Sentry, Notion, Slack, Supabase and Stripe among services that can be connected through MCP. The docs are the place to confirm setup and SDK support before committing to a particular language or CI environment.

OpenCode is open source under the MIT license, and the homepage offers free models without an account. The free-model set has limits, but the homepage does not publish its quotas. You can also bring an existing subscription, including ChatGPT Plus or Pro or GitHub Copilot, so model charges depend on the subscription you use.

OpenCode Go costs $10 per month. The pricing page lists curated open-weight models and estimated allowances of 110 Kimi K3 requests or 1,050 DeepSeek V4 Pro requests per five hours. Go Plus costs $40 per month and lists 440 Kimi K3 or 4,200 DeepSeek V4 Pro requests per five hours. These are estimates, not a fixed per-request bill. The page lists no annual billing; subscriptions can be canceled any time. Credit top-ups for Extra Usage are available, but their price is not published. Go Teams starts with Extra Usage on. Check the current allowance and top-up terms before relying on a plan for sustained production work.

OpenCode is MIT-licensed, and its source code is available on GitHub. The vendor describes it as working on your machine and supports local models, so the agent itself is not limited to a hosted interface. The product is presented as a desktop app, terminal UI, mobile app and SDK; the vendor does not present a separate self-hosted server product.

Model choice across more than 200 providers, including local models, gives teams options if they change model vendors. You can also bring an existing subscription rather than use Go. That reduces dependence on one model subscription, but does not establish that every workflow or configuration transfers cleanly. The vendor does not publish data-export details in the homepage text. Before adopting it for a team, check how conversations, settings and project changes are stored and exported, and how to preserve any agent configuration you depend on.

Before putting OpenCode in a production workflow, check free-model quotas, Go allowances, Extra Usage pricing and the behavior when a model limit is reached. The pricing page gives estimated requests per five hours for Go and Go Plus, but those figures are model-specific. The vendor does not publish a rate-limit policy, uptime commitment or support response target in the homepage text. For CI, confirm SDK language support and failure handling in the docs.

Treat permissions and sandboxing as part of deployment review. One Reddit commenter said V2 no longer asked for permission to access other directories and argued that sandboxing mattered more as a result. Other Reddit users report that V2 works fine, including with local models; another thread includes complaints about hard limits on free use. These are individual reports, not a reliability assessment. Test the exact model, permissions and workload you plan to use, and set a fallback for provider outages or exhausted allowances.