Plugin Installer

OpenCode Plugin Installer

Help users discover, evaluate, and install OpenCode plugins from the community catalog.

<workflow>

1. List Available Plugins

Run the catalog script to see what's available:

python3 ~/.config/opencode/skill/plugin-installer/scripts/list_plugins.py

Output shows name: description with path to detailed file.

2. Read Plugin Details

For relevant matches, read the full plugin file:

references/plugins/<filename>.md

Each file contains:

• Installation config (copy-paste JSON)
• Setup steps
• Features and caveats
• Links to docs

3. Recommend and Install

Show the user:

• What the plugin does
• Install snippet for opencode.json
• Any setup steps (OAuth, config files, etc.)

</workflow>

<question_tool>

Batching: Use the question tool for 2+ related questions. Single questions → plain text.

Syntax: header ≤12 chars, label 1-5 words, add "(Recommended)" to default.

When to ask: Multiple plugins match the need, or plugin has complex setup.

</question_tool>

<installation>

Installation Format

Add to plugin array in opencode.json:

{
  "plugin": [
    "package-name@version"
  ]
}

Config locations:

• Global: ~/.config/opencode/opencode.json
• Project: .opencode/opencode.json

Version Pinning

// Pin version (RECOMMENDED)
"plugin": ["package@1.2.3"]

// Auto-update on restart
"plugin": ["package@latest"]

// Locked to first install (never updates)
"plugin": ["package"]

Force update when stuck:

rm -rf ~/.cache/opencode/node_modules ~/.cache/opencode/bun.lock

</installation>

<documenting_new_plugins>

Adding New Plugins to the Catalog

When you discover a new OpenCode plugin (from web search, GitHub, npm, user mention), you MUST document it for future reference.

Step 1: Gather Plugin Information

Before creating the doc, collect:

Step 2: Create the Plugin File

Location: ~/.config/opencode/skill/plugin-installer/references/plugins/<name>.md

Naming convention: Use the npm package name with / replaced by - for scoped packages:

• opencode-foo → opencode-foo.md
• @org/opencode-bar → org-opencode-bar.md

Step 3: Use This Template

---
name: package-name-or-@scope/package-name
description: One-line description for catalog listing (max ~80 chars)
---
# Plugin Display Name

Brief description of what this plugin does and why someone would use it.

## Installation

\`\`\`jsonc
{
  "plugin": ["package-name@latest"]
}
\`\`\`

## Setup

### Prerequisites
- List any requirements (accounts, API keys, other tools)

### Configuration
Step-by-step setup instructions.

## Features
- Key feature 1
- Key feature 2

## Caveats
- Known limitations
- Compatibility notes

## Links
- [GitHub](https://github.com/...)
- [npm](https://www.npmjs.com/package/...)

Step 4: Verify the Catalog

After creating, run the listing script to confirm it appears:

python3 ~/.config/opencode/skill/plugin-installer/scripts/list_plugins.py

</documenting_new_plugins>

<frontmatter_requirements>

Important: The name MUST match the npm package name exactly (including scope if any). This is what users will copy into their opencode.json.

</frontmatter_requirements>

<example>

Documenting a New Plugin

Say you found @cooldev/opencode-metrics on npm. Create:

File: ~/.config/opencode/skill/plugin-installer/references/plugins/cooldev-opencode-metrics.md

---
name: @cooldev/opencode-metrics
description: Track AI usage metrics and export to dashboards
---
# OpenCode Metrics

Export your OpenCode usage data to external dashboards...

## Installation

\`\`\`jsonc
{
  "plugin": ["@cooldev/opencode-metrics@latest"]
}
\`\`\`

...

</example>

LLM-API-Key-Proxy

Type: External Tool (not npm plugin) Author: Mirrowel Status: Use dev/beta release for latest features

Purpose

Universal LLM Gateway - one OpenAI-compatible API endpoint for all providers. Self-hosted proxy with automatic key rotation, failover, and exclusive provider support.

Features

• Single /v1/chat/completions endpoint for all providers
• Automatic key rotation and load balancing
• Rate limit handling with intelligent cooldowns
• Escalating cooldowns per model (10s → 30s → 60s → 120s)
• OAuth support: Gemini CLI, Antigravity, Qwen Code, iFlow
• Interactive TUI for configuration
• Detailed request logging

Exclusive Providers

Providers not available elsewhere:

Installation

1. Download from GitHub Releases

• Use dev/beta release for Antigravity and latest features

2. Extract and run:

• Windows: proxy_app.exe
• macOS/Linux: chmod +x proxy_app && ./proxy_app

3. Configure credentials via TUI

From Source

git clone https://github.com/Mirrowel/LLM-API-Key-Proxy.git
cd LLM-API-Key-Proxy
python3 -m venv venv
source venv/bin/activate
pip install -r requirements.txt
python src/proxy_app/main.py

OpenCode Configuration

{
  "provider": {
    "local": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "LLM Proxy",
      "options": {
        "baseURL": "http://127.0.0.1:8000/v1",
        "apiKey": "your-PROXY_API_KEY-from-env"
      },
      "models": {
        "gemini-2.5-flash": {
          "id": "gemini/gemini-2.5-flash",
          "name": "Gemini 2.5 Flash"
        },
        "antigravity-gemini-3": {
          "id": "antigravity/gemini-3-pro-preview",
          "name": "Gemini 3 Pro (Antigravity)",
          "reasoning": true
        },
        "antigravity-claude-opus": {
          "id": "antigravity/claude-opus-4-5",
          "name": "Claude Opus 4.5 (Antigravity)",
          "reasoning": true
        },
        "gemini-cli-pro": {
          "id": "gemini_cli/gemini-2.5-pro",
          "name": "Gemini 2.5 Pro (CLI)"
        }
      }
    }
  }
}

Model Format

Important: Models must be provider/model_name:

gemini/gemini-2.5-flash          # Gemini API
openai/gpt-4o                    # OpenAI API
anthropic/claude-3-5-sonnet      # Anthropic API
openrouter/anthropic/claude-3-opus  # OpenRouter
gemini_cli/gemini-2.5-pro        # Gemini CLI (OAuth)
antigravity/gemini-3-pro-preview # Antigravity
antigravity/claude-opus-4-5      # Claude Opus 4.5
antigravity/claude-sonnet-4-5    # Claude Sonnet 4.5

Credential Setup

API Keys

Create .env file:

PROXY_API_KEY="your-secret-proxy-key"
GEMINI_API_KEY_1="key1"
GEMINI_API_KEY_2="key2"
OPENAI_API_KEY_1="sk-..."
ANTHROPIC_API_KEY_1="sk-ant-..."

OAuth (Antigravity, Gemini CLI)

python -m rotator_library.credential_tool
# Select "Add OAuth Credential" → provider
# Complete browser authentication

API Endpoints

Advanced Config

# Concurrency per provider
MAX_CONCURRENT_REQUESTS_PER_KEY_OPENAI=3

# Rotation mode
ROTATION_MODE_GEMINI=sequential  # or balanced

# Model filtering
IGNORE_MODELS_OPENAI="*-preview*"
WHITELIST_MODELS_GEMINI="gemini-2.5-pro,gemini-2.5-flash"

Caveats

• Requires running separate proxy process
• Use dev/beta release for Antigravity and Gemini 3 support
• OAuth credentials need browser authentication
• Gemini 3 models require paid-tier Google Cloud project

Links

• GitHub
• Releases

@tarquinen/opencode-dcp

Package: @tarquinen/opencode-dcp Author: Dan Tarquinen Latest Version: 1.0.4

Purpose

Dynamic Context Pruning - automatically reduces token usage by removing obsolete tool outputs from conversation history.

Features

• Deduplication: Identifies repeated tool calls, keeps only most recent output (zero LLM cost)
• Supersede Writes: Prunes write tool inputs for files that were later read (zero LLM cost)
• Prune Tool: AI can manually trigger pruning when needed
• On Idle Analysis: LLM analyzes context during idle periods to find irrelevant outputs

Session history is never modified - pruned outputs replaced with placeholder before sending to LLM.

Installation

{
  "plugin": ["@tarquinen/opencode-dcp@latest"]
}

Configuration

Config file locations:

• Global: ~/.config/opencode/dcp.jsonc (auto-created on first run)
• Custom: $OPENCODE_CONFIG_DIR/dcp.jsonc
• Project: .opencode/dcp.jsonc

Precedence: Defaults → Global → Config Dir → Project

Example Config

{
  "deduplication": {
    "enabled": true,
    "protectedTools": []
  },
  "supersedeWrites": {
    "enabled": true
  },
  "pruneTool": {
    "enabled": true
  },
  "onIdleAnalysis": {
    "enabled": false,
    "idleThresholdMs": 30000
  }
}

Protected Tools

These tools are always protected from pruning:

• task
• todowrite
• todoread
• prune
• batch

Add more via protectedTools arrays in each strategy.

Trade-offs

Prompt Caching Impact: Pruning changes message content, which invalidates cached prefixes. You lose some cache read benefits but gain larger token savings from reduced context size.

In most cases, token savings outweigh cache miss cost - especially in long sessions.

Caveats

• Restart OpenCode after config changes
• Does not modify actual session history
• On Idle Analysis uses LLM calls (has cost)

Links

• npm
• GitHub

opencode-gemini-auth

Package: opencode-gemini-auth Author: jenslys Latest Version: Check npm

Purpose

Authenticate with your Google account to use Gemini models via OAuth. Use your existing Gemini plan and quotas (including free tier) without separate API billing.

Features

• OAuth with Google account
• Use existing Gemini quotas (free tier included)
• Thinking config support:
• thinkingLevel (low/high) for Gemini 3
• thinkingBudget (token count) for Gemini 2.5
• Zero-config Google Cloud project discovery
• Auto-provisions suitable GCP project

Installation

{
  "plugin": ["opencode-gemini-auth@latest"]
}

Setup

1. Add plugin to config 2. Restart OpenCode 3. Run opencode auth login 4. Select "Google" → "OAuth with Google (Gemini CLI)" 5. Complete browser authentication

Configuration

Specify Project ID

{
  "provider": {
    "google": {
      "options": {
        "projectId": "your-specific-project-id"
      }
    }
  }
}

Gemini 3 Thinking (thinkingLevel)

{
  "provider": {
    "google": {
      "models": {
        "gemini-3-pro-preview": {
          "options": {
            "thinkingConfig": {
              "thinkingLevel": "high",
              "includeThoughts": true
            }
          }
        }
      }
    }
  }
}

Gemini 2.5 Thinking (thinkingBudget)

{
  "provider": {
    "google": {
      "models": {
        "gemini-2.5-flash": {
          "options": {
            "thinkingConfig": {
              "thinkingBudget": 8192,
              "includeThoughts": true
            }
          }
        }
      }
    }
  }
}

Troubleshooting

Manual Google Cloud Setup

If auto-provisioning fails: 1. Go to Google Cloud Console 2. Create or select a project 3. Enable Gemini for Google Cloud API (cloudaicompanion.googleapis.com) 4. Set projectId in config

Debug Mode

OPENCODE_GEMINI_DEBUG=1 opencode

Generates gemini-debug-<timestamp>.log files.

Updating

rm -rf ~/.cache/opencode/node_modules/opencode-gemini-auth
opencode

Links

• npm
• GitHub

@franlol/opencode-md-table-formatter

Package: @franlol/opencode-md-table-formatter Author: franlol Latest Version: 0.0.3

Purpose

Automatically formats markdown tables in AI responses. Correctly calculates column widths when markdown concealment mode is enabled.

Features

• Automatic table formatting after AI text completion
• Concealment mode compatible (correct widths when markdown symbols hidden)
• Alignment support: left (:---), center (:---:), right (---:)
• Nested markdown handling with multi-pass algorithm
• Code block preservation (` bold ` shows literal)
• Handles emojis, unicode, empty cells, long content
• Silent operation - no console logs, errors don't interrupt
• Validation feedback for invalid tables

Installation

{
  "plugin": ["@franlol/opencode-md-table-formatter@latest"]
}

Example

Input (unformatted):

| Feature | Description | Status |
|---|---|
| **Bold text** | Has *italic* content | ✅ Done |
| `Code` | With `**bold**` inside | ⏳ Progress |

Output (formatted):

| Feature       | Description            | Status      |
| ------------- | ---------------------- | ----------- |
| **Bold text** | Has *italic* content   | ✅ Done     |
| `Code`        | With `**bold**` inside | ⏳ Progress |

How It Works

Uses OpenCode's experimental.text.complete hook to format tables after AI finishes generating text. Strips markdown symbols for width calculation while preserving symbols inside inline code blocks.

Troubleshooting

Tables not formatting?

• Ensure plugin is in config
• Restart OpenCode
• Tables must have separator row (|---|---|)

Invalid table structure comment?

• All rows must have same number of columns
• Tables need at least 2 rows including separator

Caveats

• Only works with concealment mode enabled (OpenCode default)
• Requires Node.js >= 18.0.0 or Bun runtime

Links

• npm
• GitHub

opencode-openai-codex-auth

Package: opencode-openai-codex-auth Author: numman-ali Latest Version: 4.2.0

Purpose

Use your ChatGPT Plus/Pro subscription instead of OpenAI API credits. OAuth authentication via ChatGPT using the same method as the official Codex CLI.

Features

• OAuth authentication via ChatGPT Plus/Pro
• 22 pre-configured model variants (GPT 5.2, GPT 5.1 Codex, Codex Max, Codex Mini)
• Reasoning levels: none/low/medium/high/xhigh
• Prompt caching support
• Full image input support for all models
• Auto-refreshing tokens
• CODEX_MODE bridge prompt with Task tool & MCP awareness

Installation

{
  "plugin": ["opencode-openai-codex-auth@4.2.0"]
}

Full Configuration (Recommended)

{
  "$schema": "https://opencode.ai/config.json",
  "plugin": ["opencode-openai-codex-auth@4.2.0"],
  "provider": {
    "openai": {
      "options": {
        "reasoningEffort": "medium",
        "reasoningSummary": "auto",
        "textVerbosity": "medium",
        "include": ["reasoning.encrypted_content"],
        "store": false
      },
      "models": {
        "gpt-5.2-high": {
          "name": "GPT 5.2 High (OAuth)",
          "limit": { "context": 272000, "output": 128000 },
          "options": {
            "reasoningEffort": "high",
            "reasoningSummary": "detailed"
          }
        }
      }
    }
  }
}

Setup

1. Add plugin to config 2. Restart OpenCode 3. Run opencode auth login 4. Select "OpenAI" → "ChatGPT Plus/Pro (Codex Subscription)" 5. Complete browser OAuth flow

Available Models

Caveats

• Personal use only - not for commercial/multi-user applications
• Must use full config from config/full-opencode.json in the repo
• GPT 5 models can be temperamental - stick to presets
• Stop Codex CLI if running (both use port 1455)

Updating

Change version number in config and restart. If stuck on old version:

rm -rf ~/.cache/opencode/node_modules ~/.cache/opencode/bun.lock

Troubleshooting

• 401 Unauthorized: Run opencode auth login again
• Model not found: Add openai/ prefix (e.g., --model=openai/gpt-5.2-codex-high)

Links

• npm
• GitHub
• Documentation

@ramtinj95/opencode-tokenscope

Package: @ramtinj95/opencode-tokenscope Author: ramtinJ95 Latest Version: Check npm

Purpose

Comprehensive token usage analysis and cost tracking for OpenCode sessions. Track and optimize token usage across system prompts, user messages, tool outputs, and more.

Features

• Token breakdown by category (system, user, tools, assistant)
• Tool usage breakdown with call counts and percentages
• Top token contributors identification
• Cache efficiency analysis (hit rate, cost savings)
• Cost estimation with model-specific pricing
• Subagent (Task tool) cost tracking across child sessions
• Context breakdown (system prompt, tool definitions, project tree)
• Visual ASCII charts

Installation

{
  "plugin": ["@ramtinj95/opencode-tokenscope@latest"]
}

Usage

Adds a tokenscope tool that can be called by the AI or requested by user.

Example output:

═══════════════════════════════════════════════════════════════
Token Analysis: Session ses_abc123
Model: claude-opus-4-5
═══════════════════════════════════════════════════════════════

TOKEN BREAKDOWN BY CATEGORY
───────────────────────────────────────────────────────────────
Input Categories:
  SYSTEM    ██████████████░░░░░░░░░░░░░░░░    45.8% (22,367)
  USER      ░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░        0.8% (375)
  TOOLS     ████████████████░░░░░░░░░░░░░░    53.5% (26,146)

CACHE EFFICIENCY
───────────────────────────────────────────────────────────────
  Cache Hit Rate:      86.2%
  Cost Savings:        $1.44  (77.6% reduction)

SUBAGENT COSTS (4 child sessions, 23 API calls)
───────────────────────────────────────────────────────────────
  Subagent Total:      $1.39  (566,474 tokens)

TOTAL:                 $1.96  (942,160 tokens, 38 calls)

Configuration

Optional config to hide sections:

{
  "showBreakdown": true,
  "showToolUsage": true,
  "showCacheEfficiency": true,
  "showSubagents": true,
  "showCost": true
}

Updating

opencode --update

The --update flag skips dependency installation for faster updates.

Caveats

• Token counts are estimated using tokenizer analysis
• Cost estimates use standard API pricing from models.json
• Free/open models marked with zero pricing

Links

• npm
• GitHub

@openhax/codex

Package: @openhax/codex Author: openhax Latest Version: 0.4.4

Purpose

Alternative OpenAI Codex OAuth plugin for OpenCode. Use ChatGPT Plus/Pro subscription for Codex models.

Installation

{
  "plugin": ["@openhax/codex@latest"]
}

Setup

1. Add plugin to config 2. Restart OpenCode 3. Run opencode auth login 4. Select OpenAI OAuth option 5. Complete browser authentication

Comparison

This is an alternative to opencode-openai-codex-auth. Choose based on:

• Active maintenance
• Feature set you need
• Model support

Check both READMEs for current feature comparison.

Links

• npm

@howaboua/pickle-thinker

Package: @howaboua/pickle-thinker Author: howaboua Latest Version: 0.4.0

Purpose

Reliably injects the magic keyword Ultrathink so GLM-4.6 / Big Pickle models consistently enter "thinking mode".

Features

• Forces Ultrathink as first token in every user message
• Injects synthetic "continue + ultrathink" after tool results
• Appends Ultrathink block to tool outputs as safety net
• Re-injects on session compaction (when conversations are summarized)
• Two modes: lite (prefix only) or tool (redundant injection - recommended)

Installation

{
  "plugin": ["@howaboua/pickle-thinker@0.4.0"]
}

Configuration

Config file locations:

• Global: ~/.config/opencode/pickle-thinker.jsonc (auto-created)
• Project: .opencode/pickle-thinker.jsonc

Example Config

{
  "mode": "tool",  // "lite" or "tool" (default: tool)
  "prefix": "Ultrathink"  // Can customize but Ultrathink required for GLM-4.6
}

How It Works

Only activates for target models (GLM-4.6 + Big Pickle):

1. User messages: Forced to start with Ultrathink as dedicated leading text part 2. After tool results: Synthetic user "continue + ultrathink" message injected 3. Tool outputs/errors: Appended with Ultrathink block as extra safety 4. Session compaction: Re-injects instructions to maintain thinking mode

Duplication is intentional - missing Ultrathink is the bug.

Caveats

• Only activates for GLM-4.6 and Big Pickle models
• tool mode increases turns/tokens (may affect subscription usage)
• Ultrathink keyword required - custom prefix still ensures it's first

Acknowledgments

Based on the Dynamic Context Pruning plugin architecture by Dan Tarquinen.

Links

• npm