OpenCode Plugin for MorphLLM

2026-02-08

Background

I was burning through my GitHub Copilot quota on trivial stuff. "Update the README" doesn't need Opus, but every prompt ate the same monthly allowance. Mid-month I'd be rationing requests or settling for weaker models.

MorphLLM has a classifier that tags prompts as easy, medium, or hard. I built opencode-morphllm to route prompts to appropriate models based on complexity.

The Router

The plugin hooks OpenCode's chat.message event:

// src/morph/router.ts
export function createModelRouterHook() {
  return {
    'chat.message': async (input, output) => {
      if (!MORPH_ROUTER_ENABLED) return;

      const promptText = extractPromptText(output.parts);
      const classification = await getMorphClient()
        .routers.raw.classify({ input: promptText });
      
      const chosen = pickModelForDifficulty(classification?.difficulty);
      input.model.providerID = chosen.providerID || input.model.providerID;
      input.model.modelID = chosen.modelID || input.model.modelID;
    },
  };
}

Morph returns easy, medium, or hard. The pickModelForDifficulty maps those to configured models:

function pickModelForDifficulty(difficulty?: string) {
  switch (String(difficulty).toLowerCase()) {
    case 'easy':   return parseModel(MORPH_MODEL_EASY);
    case 'medium': return parseModel(MORPH_MODEL_MEDIUM);
    case 'hard':   return parseModel(MORPH_MODEL_HARD);
    default:       return parseModel(MORPH_MODEL_DEFAULT);
  }
}

There's also a caching mode where it sticks to the first model chosen per session—useful if your provider charges less for cached prompts.

MCP Tools

The plugin injects Morph's MCP tools through the config hook:

// src/morph/mcps.ts
export function createBuiltinMcps() {
  return {
    morph_mcp: {
      type: 'local',
      command: ['npx', '-y', '@morphllm/morphmcp'],
      environment: {
        MORPH_API_KEY: API_KEY,
        ENABLED_TOOLS: 'edit_file,warpgrep_codebase_search',
      },
      enabled: true,
    },
  };
}

edit_file uses Morph's Fast Apply. Instead of diff matching, you provide the original code plus a snippet with // ... existing code ... placeholders. It merges semantically, preserving imports and structure. Morph claims ~98% accuracy at 10,500+ tokens/sec.

warpgrep_codebase_search runs as a subagent in its own context. Instead of dumping raw grep results into your main conversation, it performs multi-step searches (up to 24 tool calls across 4 turns) and returns relevant code. Keeps your main context clean—no embedding index required.

My Setup

{
  "MORPH_API_KEY": "key",
  "MORPH_ROUTER_CONFIGS": {
    "MORPH_MODEL_EASY": "github-copilot/gpt-5-mini",
    "MORPH_MODEL_MEDIUM": "opencode/kimi-k2.5-free",
    "MORPH_MODEL_HARD": "opencode/kimi-k2.5-free",
    "MORPH_ROUTER_ENABLED": true
  }
}

Easy stuff goes to Copilot's cheaper tier. Medium and hard go to free Kimi models. Premium usage dropped 70-80%. Classification adds ~50-100ms.

Install by adding to ~/.config/opencode/opencode.json:

{
  "plugins": ["opencode-morphllm"]
}

OpenCode handles npm installation.

Continuation

Some things I want to explore: