OpenCode is an MIT-licensed coding agent that pioneered the skill-driven execution model. Instead of treating every request as a fresh prompt and hoping the model improvises the right steps, OpenCode packages workflows into reusable SKILL.md files. The result is an agent that behaves less like an unpredictable chatbot and more like a set of documented, repeatable procedures — which is exactly what teams want when they put an agent anywhere near production code.
This deep dive covers how the skill model works, the slash-command surface, the agent lifecycle, and where OpenCode fits relative to commercial agents.
The Skill Execution Model
Every task in OpenCode maps to a skill. Skills live in a /skills directory as SKILL.md files, each containing frontmatter (name, description, allowed tools) and step-by-step instructions. When an incoming request matches a skill’s description, OpenCode loads that skill and follows its workflow instead of free-forming a plan.
This matters for two reasons. First, reproducibility: the same request runs the same steps every time, so you get consistent output instead of a different approach on every invocation. Second, auditability: because the workflow is a plain Markdown file in your repo, anyone can read exactly what the agent will do before it does it. There is no opaque system prompt to reverse-engineer — the behavior is the file.
Slash Commands
OpenCode exposes skills through slash commands, giving you a predictable surface instead of hoping natural language routes correctly:
/plan— break a task into steps before touching code/spec— turn a rough idea into a concrete specification/review— critique a diff against style and correctness rules/test— generate and run tests for changed code/ship— commit, and prepare a PR with a summary/debug— reproduce and isolate a failure
Each command loads a specific skill with its own steps, tool permissions, and exit criteria. You can add your own commands by dropping a new SKILL.md in the skills directory — no plugin registry, no rebuild.
Agent Lifecycle
OpenCode’s loop is deliberately explicit. It maps the request to a skill, loads the skill instructions, executes each step, verifies outputs against the skill’s exit criteria, and reports results. Because verification is a first-class step rather than an afterthought, the agent catches its own half-finished work — a skill that says “run the tests and confirm they pass” will not report success on a red suite.
That structured lifecycle is the whole point: consistent output quality comes from consistent process, not from a bigger model.
Configuring OpenCode With AGENTS.md
Like most terminal agents, OpenCode reads project conventions from AGENTS.md — build commands, code style, directories to avoid, and review rules. Pairing a good AGENTS.md with a set of skills is the highest-leverage setup for any team, because the two files together encode both what your project is and how the agent should work in it. Start from our complete AGENTS.md guide if you have not written one yet.
Security Considerations
Skills declare which tools they are allowed to use, which gives you per-workflow permission scoping instead of one blanket grant. That is a genuine security advantage, but self-hosting an open-source agent also means the sandboxing and credential handling are your responsibility, not a vendor’s. Before you point OpenCode at a real repo with real secrets, run through our coding agent security checklist — the section on credential exposure and write-scope escapes is directly relevant to any agent that can edit files and run commands.
How OpenCode Compares
Against Hermes. OpenCode is simpler and more focused. It does not ship mixture-of-agents, cron jobs, persistent memory, or a console dashboard. But its skill execution model is more explicit and easier to customize. For teams that want transparent, hackable workflows over a batteries-included platform, OpenCode is the more compelling starting point.
Against commercial agents. The trade is the usual open-source calculus: you give up the polished gateway, hosted auth, and vendor support that something like Claude Code provides, and in return you get source you can read, fork, and self-host. Our open-source vs commercial coding agents guide works through that decision in detail, and the coding agent features comparison puts the capability matrix side by side.
Model flexibility. Because OpenCode is provider-agnostic, you can point it at whichever model API you want — including cheap models for bulk work and stronger ones for hard reasoning. That routing freedom is a real cost lever; see the coding agent pricing guide for how that plays out on the bill.
Who Should Use OpenCode
Reach for OpenCode if you value transparency, want to encode your team’s workflows as version-controlled skills, and are comfortable owning your own hosting and security. Look at a commercial agent instead if you need an enterprise gateway, hosted authentication, and a support contract out of the box. For most engineers who like reading the code they run, the skill-driven model is worth trying — and because it is MIT-licensed, the only cost of evaluating it is your time. To place it against the rest of the field, the best coding agents decision guide maps tools to team size and workload.
Why pay $20/mo for ChatGPT, $20/mo for Claude, $20/mo for Gemini? aiFiesta gives you all of them and more for just $12/mo. One subscription, every answer.