1. Install OpenCode CLI

OpenCode is the agent harness that runs everything. Install globally:

npm install -g opencode

Configuration lives at ~/.config/opencode/opencode.jsonc. This file controls plugins, MCP servers, and provider settings.

OpenCode on GitHub ↗ · opencode.ai ↗

2. Configure Plugins

Add these four plugins. Each provides a specific optimization layer. Paste this into your opencode.jsonc:

{
  "plugin": [
    "@mohak34/opencode-notifier@latest",
    "@tarquinen/opencode-dcp@latest",
    "@franlol/opencode-md-table-formatter@latest",
    "superpowers@git+https://github.com/obra/superpowers.git"
  ]
}

3. Add MCP Servers

MCP servers provide structured tool access — browser automation, GitHub API, web search, Next.js diagnostics. They return parsed, structured output instead of raw HTML/logs, saving thousands of tokens per invocation.

"mcp": {
  "chrome-devtools": {
    "type": "local",
    "command": ["npx", "-y", "chrome-devtools-mcp@latest"]
  },
  "github": {
    "type": "remote",
    "url": "https://api.githubcopilot.com/mcp/",
    "headers": {
      "Authorization": "Bearer {file:~/.config/opencode/.secrets/github-pat}"
    }
  },
  "next-devtools": {
    "type": "local",
    "command": ["npx", "-y", "next-devtools-mcp@latest"]
  },
  "exa": {
    "type": "remote",
    "url": "https://mcp.exa.ai/mcp"
  }
}

Store your GitHub Personal Access Token at ~/.config/opencode/.secrets/github-pat. The {file:...} directive tells OpenCode to read it from disk — never put tokens directly in config files.

4. Superpowers Skills (14 built-in)

The Superpowers plugin includes 14 skills that enforce a disciplined workflow. They are auto-loaded — no manual install needed once the plugin is configured in Step 2.

5. Community Skills (53 via agentskill.sh)

Install via /learn @author/skill-name in OpenCode. Each skill listed with its agentskill.sh ↗ link. Some links may 404 if the author or skill name differs — search agentskill.sh directly to find the exact listing.

6. Provider Setup

Models cached locally in SQLite, refreshed every 15 minutes. Configure your API key once.

DeepSeek V4 Pro delivered Ignite for $1.77 with 75% permanent discount and prompt caching (99.7% savings on cached prefixes). DeepSeek API Pricing Docs ↗

7. Token Efficiency Explained

This setup reduced Ignite's build cost from an estimated $50-115 (on GPT-4o or similar premium model) to $1.77. Here's how each layer contributes:

8. Putting It Together — Sample Workflow

Here's a real workflow using this setup to build a new project:

1. Start a new project — Open OpenCode in your project directory
2. Brainstorm — The Superpowers skill kicks in automatically. Answer questions one at a time until the design is approved.
3. Write a plan — A bite-sized implementation plan (17 tasks) is generated with exact file paths, code, and verification steps.
4. Dispatch subagents — One fresh subagent per task. Each starts with only the task brief — zero context from previous tasks.
5. Review each task — After each subagent finishes, a reviewer checks spec compliance + code quality.
6. Push to GitHub — When all tasks pass review, commit and push. Done.

For Ignite specifically, this workflow produced 38 commits across 17 tasks in a single session — from empty directory to shipped .app bundle — for $1.77.

Pro Tips

• Use cheap models for mechanical tasks. Save DeepSeek V4 Pro for architecture, review, and debugging. Use DeepSeek Flash for single-file edits and boilerplate.
• Keep the pipeline tight. Don't skip brainstorm (the most expensive waste is building the wrong thing). Don't skip plan (ambiguity causes subagent re-dispatches).
• Watch your context. If live tokens exceed 30%, DCP needs tuning. If they grow fast, your agents are holding too much state.
• Pin your versions. Plugins, MCPs, and skills all evolve. Pin them in your config for reproducibility across your team.
• Install all skills upfront. The 67 skills are cached as a system prompt prefix — installing them costs nothing in ongoing token usage since they sit in cache.