OpenCode

OpenCode is an open-source AI coding agent from Anomaly (the team behind SST). It runs in the terminal, IDE, or desktop, connects to model providers via Models.dev, and stores zero code or context data server-side. MIT-licensed and fully self-hostable, you bring your own model key (Claude, GPT, Gemini, local Llama) and pay nothing for the orchestrator itself, or upgrade to OpenCode Go at $5/month or curated OpenCode Zen for managed coding-tuned models.

What OpenCode does:

• •Open-source and MIT-licensed the entire agent, TUI, and IDE extensions are MIT on GitHub at github.com/sst/opencode (156,000+ stars, 850+ contributors, 6.5M monthly developers), so you can self-host the binary on a corporate workstation, audit every line, or fork freely.
• •Self-hosting and zero data retention OpenCode stores no code or context server-side, which means regulated and air-gapped teams can run it on their own machines without leaking source.
• •75+ model providers via Models.dev swap between Claude, GPT-5, Gemini, Llama, DeepSeek, Qwen, or any local Ollama model with one config change; no provider lock-in.
• •Native terminal TUI a fast keyboard-first text user interface optimised for terminal-only workflows; no Electron app, no browser tab, just a terminal.
• •Parallel multi-session work run multiple coding agents simultaneously on the same project (different branches or features) and share session URLs for debugging or pair work.
• •LSP integration OpenCode automatically loads the right Language Server Protocol per file, which gives the agent type information, definitions, and references for sharper edits.
• •MCP support bring any MCP server into OpenCode for filesystem, database, or third-party tool access, and expose OpenCode itself to other MCP clients.
• •GitHub Copilot and ChatGPT Plus/Pro auth reuse your existing Copilot or ChatGPT subscription for inference instead of paying per-token; no extra API key management.
• •OpenCode Zen (paid) -- a curated, benchmarked model gateway optimised for agentic coding tasks; quality control on which checkpoints work reliably for autonomous code work.
• •CLI-first and desktop app the CLI is the primary interface, with a desktop app on macOS, Windows, and Linux (beta) that wraps the same agent for less terminal-comfortable users.

Pricing:

• •Self-hosted with your own keys $0/month: free forever, MIT-licensed; you pay only for whatever LLM API you connect (or run a local model on your own GPU).
• •OpenCode Go $5/month: subsidised access to open-source-friendly models for users who do not want to manage API keys themselves.
• •OpenCode Zen custom pricing: curated, benchmarked premium models tuned for agentic coding; pay-as-you-go credits managed by the OpenCode team.
• •Enterprise custom contract pricing: priority support, security review, on-premise installation help.

Limitations:

• •Setup friction as of 2026, getting OpenCode configured with local models or non-default providers requires more technical knowledge than Cursor or Copilot, so non-technical users will struggle on the first install.
• •No native browser extension OpenCode is terminal-first, so workflows that depend on browser-context coding (Replit-style) need a different tool.
• •Beta desktop app on Linux Linux desktop is still beta in 2026 (terminal CLI is stable on Linux); rough edges and occasional crashes happen there.
• •No public REST API the agent is interactive-first; teams that want to call OpenCode programmatically rely on MCP integration or shelling out to the CLI rather than a hosted API.