# OpenCode

OpenCode is a standalone agentic coding harness — a terminal-and-IDE agent that uses tool calls to read, edit, and run code against a project. It positions itself as a more configurable alternative to vendor-locked tools (Claude Code, Codex) and ships first-class support for pointing at any OpenAI-compatible endpoint, including locally-served models via [[toolbox/llama-cpp|llama.cpp]].

## What it does

A long-lived agent loop with a stable tool surface (file read/write, shell, search, edit), driven by an LLM accessed through a configurable provider. The differentiating choices are:

- **Provider-agnostic.** Configured via a JSON file with provider URL, model name, and per-provider tweaks. Local llama.cpp servers, OpenAI, Anthropic, Groq, and the bundled ZEN model set are all first-class.
- **AGENTS.md as the project contract.** OpenCode reads a project-level `AGENTS.md` (analogous to Cursor Rules or `CLAUDE.md`) for context, conventions, and rules. This is the file that holds nonstandard filename-prefix requirements, typing constraints, comment styles, etc.
- **Bundled ZEN model set.** Vetted hosted models on a hosted subscription, including free tiers. Useful for users who want OpenCode-the-harness without separately picking a model and provider.

For Vyacheslav's [[local-llm-16gb-vram-tests|coding-agent tests on local 16 GB models]], OpenCode + llama.cpp was the rig. He picked it over Cline, Kilo Kode, Roo Code, and Zed for "stability and simplicity of setup."

## How it pairs with local models

The pairing requires three things on the llama.cpp side and one on the OpenCode side:

- `--jinja` flag on `llama-server` — required for the chat template to render tool calls in the format OpenCode expects. Without it, the model never sees the tool list.
- A quantization that the specific model architecture survives — see [[llm-quantization]].
- Enough VRAM headroom for both model and KV cache — see [[kv-cache-sizing]] and [[moe-cpu-offload]].

On the OpenCode side, provider config points at `http://localhost:8080/v1` (or wherever llama.cpp is bound) and sets the model name. For Qwen 3-family models, the OpenCode config also needs `extra_body.chat_template_kwargs.enable_thinking: false` to actually disable thinking mode — the `--reasoning off` flag at the server level is not enough (see [[thinking-mode-rule-erosion]]).

## Caveats from real use

- The "Thinking…" status indicator in OpenCode's UI is a generic in-flight indicator, not evidence the model is generating thinking tokens. Don't take it as confirmation that reasoning mode is on.
- Skills designed for large hosted models (Context 7, Superpowers, anything that injects a lot of system prompt) overwhelm small local models' context. Vyacheslav recommends "caveman" or no skills for local-model setups.
- The system prompt + skills + MCP context all consume context window before the user's first message. This compounds with small models' already-tight context budgets.

## Why it shows up in the wiki

OpenCode is the harness against which several findings get measured: [[thinking-mode-rule-erosion]] was discovered by running through OpenCode with `AGENTS.md` rules; the MoE offload numbers in [[moe-cpu-offload]] are with OpenCode driving the inference. It's also one of the harnesses in [[toolbox/cli-anything|cli-anything's]] cross-harness skill distribution.

## Related

- [[toolbox/llama-cpp|llama.cpp]] — the inference engine OpenCode points at for local models
- [[lucumr-local-models]] — Ronacher's critique of local-LLM UX gaps; OpenCode is a candidate to fill some of them
- [[toolbox/cli-anything]] — cross-harness skill packaging that includes OpenCode
- [[toolbox/peon-ping]] — has hooks for OpenCode among other harnesses

Repo: https://github.com/sst/opencode · TypeScript · MIT