LLM-grounded web search plugin for OpenCode that returns structured grounding analysis: citation-ranked sources, coverage percentage, ungrounded claims (potential hallucination risk), and search query trace.

This plugin exposes opencode_gemini_search as a custom tool, so your agent can perform grounded web search with rich analysis instead of just raw text + citations.

Features

• opencode_gemini_search tool backed by Gemini's Grounding with Google Search
• Ranked sources table sorted by citation frequency with confidence scores
• Coverage percentage — what fraction of the answer is actually grounded
• Ungrounded claims — sentences with no grounding (potential hallucination risk)
• Search query trace — what sub-queries the model ran internally
• Full sources list with titles and URIs

Example output:

## Answer

Quantum computing is advancing rapidly[1][2]. Researchers at IBM achieved new milestones[3]. Some ungrounded claim here.

---

### Sources (ranked by citation frequency)
| # | Source | Domain | Citations | Confidence |
|---|--------|--------|-----------|------------|
| 1 | IBM Research | ibm.com | 3 | high |
| 2 | Google DeepMind | deepmind.google | 2 | medium |

### Coverage
- **Grounded:** 87.0%
- **Ungrounded:** 13.0%

### Ungrounded Claims (potential hallucination risk)
1. "Some ungrounded claim here" — no grounding

### Search Queries
- "quantum computing breakthroughs 2026"
- "latest quantum hardware developments"

### Sources List
[1] IBM Research - https://research.ibm.com/quantum
[2] Google DeepMind - https://deepmind.google/quantum

Installation

Add gemini-ground-search to your ~/.config/opencode/opencode.json.

IMPORTANT: Put gemini-ground-search LAST in the plugin list to avoid impacting other plugins' auth process, and disable the plugin before starting any auth process.

{
  "$schema": "https://opencode.ai/config.json",
  "plugin": [
    "...other plugins",
    "[email protected]"
  ]
}

OpenCode does not upgrade plugins automatically, so you need to pin the version once the plugin is upgraded.

As long as the plugin is enabled and the provider auth is configured, any OpenCode agent that can use tools will be able to call opencode_gemini_search when it needs grounded web search.

Configure web search

Log in with opencode auth login first.

Authenticate with a Google API key.

Set a opencode_gemini_search model in your OpenCode config (optional, defaults to gemini-2.5-flash):

{
  "$schema": "https://opencode.ai/config.json",
  "provider": {
    "google": {
      "options": {
        "opencode_gemini_search": {
          "model": "gemini-2.5-flash"
        }
      }
    }
  }
}

If auth or model config is missing, opencode_gemini_search throws an error and OpenCode will display the message.

Development

This repository uses TypeScript and esbuild.

# Install dependencies
npm install

# Run typecheck
npm run typecheck

# Run tests
npm run test

# Build
npm run build

When testing the plugin against a globally installed opencode CLI during development, you can point OpenCode at a local checkout using a file:// URL in your opencode.json:

{
  "$schema": "https://opencode.ai/config.json",
  "plugin": ["file:///path/to/gemini-ground-search/index.ts"]
}

Contributions and feedback are welcome.