OpenCode – Open source AI coding agent

Original Article ↗ Hacker News Discussion ↗

OpenCode, an open‑source terminal-based AI coding agent, is emerging as a popular alternative to proprietary tools like Claude Code, Gemini CLI, and Codex because it supports multiple model providers, subagents with different models, MCP tooling, and remote/web-based workflows. Commenters praise its flexibility, extensibility, and productivity benefits, but raise serious concerns about its heavy resource use, fast‑but‑buggy release cycle, default network behavior (including use of hosted web UI and “free” proxy models), and weak security/sandboxing defaults. A recurring theme is the growing tension between vendor‑locked, subsidized tools from major AI labs and more “model‑neutral” harnesses like OpenCode, with many developers willing to trade polish for control, privacy, and the ability to plug in local or cheaper models.

Overall sentiment

  • Many commenters like OpenCode as a flexible, open‑source “coding agent harness,” but a sizable group finds it buggy, resource‑hungry, and immature.
  • It’s widely used with multiple providers (OpenAI, Claude via API, Gemini, GLM, Kimi, Qwen, local models) and valued for sub‑agents, skills, and MCP integration.
  • Some say it has replaced other tools (Aider, Codex, Cursor‑style tools) for them; others have gone back to vendor tools (Claude Code, Codex, Gemini CLI).

Anthropic / Claude Code conflict & pricing

  • Claude subscription tokens are contractually restricted to Anthropic’s own tools; OpenCode is barred from using subscription auth, but can still use the Claude API at per‑token pricing.
  • Multiple users report Anthropic API costs as prohibitively high for heavy development; they see pricing as a way to push people into subsidized, lock‑in‑heavy subscriptions.
  • Some view OpenCode’s removal of first‑party Claude API integration as a principled response; others call it a “tantrum” and note API support was partially reintroduced via third‑party plugins.

Models, sub‑agents, and workflow

  • Strong appreciation for:
    • Assigning different models to sub‑agents (cheap “worker” models vs expensive planner/reviewer models).
    • Being able to switch providers easily (OpenAI, GLM, Kimi, Qwen, etc.) and use hosted open‑weights via OpenCode’s own plans.
    • MCP support and LSP integration for richer tool access and code understanding.
  • Some report large‑project struggles with certain models (e.g., GPT 5.4 “falls apart” on big repos).

Telemetry, privacy, and “local” claims

  • Serious concern that:
    • Web UI is proxied through opencode.ai rather than purely local.
    • A “small model” / Zen fallback previously sent prompts (e.g., for titles) to third‑party models without users realizing.
    • Config and plugins can be fetched from the web, raising RCE/privacy worries.
  • Maintainers and other commenters counter that:
    • The linked “telemetry” route is for serving the WebUI, not proxying all traffic.
    • Recent code paths now fall back to the main model when no small model is configured.
  • Behavior and defaults appear to have changed over time; several people mark the situation as trust‑eroding and “unclear” without careful version inspection or MITM tests.

Performance, stability, and architecture

  • Critiques:
    • Large, complex TypeScript codebase; often >1 GB RAM for a TUI.
    • Frequent releases, regressions, and memory leaks; some users abandoned it over instability.
    • TUI feels “overbearing” and not very Unix‑like; issues on some Linux setups.
  • Praise:
    • Very fast shipping of new features and fixes.
    • Good multi‑backend architecture (server, WebUI, desktop, VS Code integration).
    • Strong extensibility via plugins, skills, and agents.

Security & sandboxing

  • OpenCode has no built‑in sandbox; many users wrap it in containers, bubblewrap, or tools like nono, and bind‑mount or copy only the project directory.
  • Several argue any serious use should isolate agents from the full filesystem and credentials; others note this is true for all coding agents, not just OpenCode.

Comparisons & alternatives

  • Frequently compared tools: Claude Code (CLI and GUI), Codex, Gemini CLI, Aider, Crush, Pi, various Rust/Go‑based agents.
  • Codex and some Rust/Go tools are praised for far better performance and lower resource use.
  • Pi and others are praised for simplicity, small prompts, and being easier to extend; OpenCode is seen as more feature‑rich but heavier and more chaotic.