opencode

Harness layer — the agent program, not a process framework and not an execution runtime. A harness loads a framework’s skills/commands and drives the model’s tool-call loop; it performs no SDLC stage and connects to the ontology only through the execution primitives it enables: as patterns. Frameworks runs_on: it; runtimes runs: it.

opencode (Anomaly — the team formerly SST/Serverless Stack; MIT, open source) is a terminal-first AI coding harness whose defining trait is model-agnosticism: unlike claude-code (Anthropic models only), it drives Claude, GPT, Gemini, Bedrock, Azure, Groq, OpenRouter, and its own curated “OpenCode Zen” model set behind one agent. It is otherwise batteries-included — first-class sub-agents, MCP, a native plugin+hook system, plan/build modes, and a permissions framework — making it, alongside Claude Code, one of the two full-primitive terminal harnesses here (and the opposite pole from pi’s minimal core). It ships a TUI (native), a desktop app, and an IDE extension.

What it provides (primitives, not process)

• Sub-agents — built-in build (full tool access) and plan (read-only) agents, plus General / Explore / a Scout research agent, and background subagents.
• MCP — MCP servers supported.
• Hooks — a native plugin hook system (tool.execute.before / .after, session.idle, …) that can intercept, block, or modify tool calls.
• Slash-commands & skills — custom commands; configurable “skills,” rules, and custom tools; AGENTS.md for project config.
• Permissions & plan mode — a permissions/policies framework (e.g. the plan agent set to “ask” before edits/bash); Plan vs Build modes.
• Session persistence — session management with undo/redo and shareable web links.
• Model access — multi-provider (Anthropic/OpenAI/Google/Bedrock/Azure/Groq/OpenRouter/…) + OpenCode Zen.

Interaction surface

Terminal TUI (native), a desktop app (beta), and an IDE extension; sessions shareable via web links. Native surface is the terminal — hence subtype: terminal.

Harness profile

Capability	opencode
Skill / command format	AGENTS.md · custom slash-commands · configurable skills · custom tools
Sub-agents	first-class — build / plan / explore / scout + background subagents
MCP	yes
Hooks / extensibility	native plugin + hook system; SDK; community registry
Permissions & guardrails	permissions/policies framework + Plan vs Build modes
Model access	model-agnostic — many providers + OpenCode Zen curated set
Memory & context files	AGENTS.md; session undo/redo + shareable state
Interaction surface	terminal TUI (primary) · desktop app · IDE extension
Distribution / license	Anomaly (ex-SST), MIT / open source

Distinctive contribution

opencode is the wiki’s model-agnostic, open-source terminal harness — the demonstration that the full primitive set (pattern-fresh-context-subagents, pattern-edit-guardrails, pattern-session-handoff, pattern-context-engineering) is not unique to Claude Code, and that a framework’s skills can run on the same primitives against any model provider. Its native plugin+hook system and multi-surface reach (TUI/desktop/IDE) make it the batteries-included counterpart to pi’s deliberately minimal core, and the open counterpart to claude-code’s Anthropic-hosted one.

Patterns provided

• pattern-fresh-context-subagents — the build/plan/explore/scout agents (and background subagents) each run with their own context and scoped tools.
• pattern-edit-guardrails — the permissions/policies framework + the read-only plan agent + hooks that can block a tool call are the mutation-gating substrate.
• pattern-session-handoff — session management with undo/redo, resume, and shareable links carries working context across a boundary.
• pattern-context-engineering — AGENTS.md + configurable skills + MCP assemble the right context per turn.

Runs / spawned by (backlinks)

• Frameworks that runs_on: it — gsd · addy-agent-skills · compound-engineering · gstack · nano-spec · superpowers (each officially names OpenCode among supported agents).
• Runtimes that runs: it — sandcastle (agent provider).

See Also

• claude-code · pi — the other terminal harnesses; claude-code is the proprietary/Anthropic-models pole, pi the minimal-core pole, opencode the model-agnostic/batteries-included pole. Cross-harness comparison: harness/index.md.
• sandcastle — the runtime that spawns this harness.
• pattern-fresh-context-subagents · pattern-edit-guardrails · pattern-session-handoff · pattern-context-engineering — the patterns it provides as primitives.