OpenCode Skills Plugin

Bring Anthropic's Agent Skills Specification (v1.0) to OpenCode. This plugin automatically discovers and registers skills as dynamic tools, enabling the Agent to leverage specialized knowledge, workflows, and bundled resources.

Features

• ✅ Auto-discovery - Scans project, home, and config directories for skills
• ✅ Spec compliance - Validates against Anthropic's Skills Specification v1.0
• ✅ Dynamic tools - Each skill becomes a skills_{{name}} tool
• ✅ Path resolution - Base directory context for relative file paths
• ✅ Nested skills - Supports hierarchical skill organization
• ✅ Graceful errors - Invalid skills skipped with helpful messages

Requirements

• OpenCode SDK ≥ 1.0.126 - Required for agent context preservation and noReply message insertion pattern

Installation

Add to your opencode.json or ~/.config/opencode/opencode.json:

{
  "plugin": ["opencode-skills"]
}

OpenCode auto-installs plugins on startup.

Version Pinning

Pin to a specific version:

{
  "plugin": ["opencode-skills@x.y.z"]
}

Plugin Updates

Check installed version:

cat ~/.cache/opencode/node_modules/opencode-skills/package.json | grep version

Force update to latest:

rm -rf ~/.cache/opencode

Then restart OpenCode.

Skill Discovery

The plugin scans these locations (lowest to highest priority):

1. ~/.config/opencode/skills/ - XDG config location (or $XDG_CONFIG_HOME/opencode/skills/)
2. ~/.opencode/skills/ - Global skills (all projects)
3. $OPENCODE_CONFIG_DIR/skills/ (if set) - Custom OpenCode config directory (higher priority than global, lower than project-local)
4. .opencode/skills/ - Project-local skills (overrides all other locations)

All locations are merged. If duplicate skill names exist, the project-local version takes precedence and a warning is logged.

Quick Start

1. Create a Skill

mkdir -p .opencode/skills/my-skill

.opencode/skills/my-skill/SKILL.md:

---
name: my-skill
description: A custom skill that helps with specific tasks in my project
license: MIT
---

# My Custom Skill

This skill helps you accomplish specific tasks.

## Instructions

1. First, do this
2. Then, do that
3. Finally, verify the results

You can reference supporting files like `scripts/helper.py` or `references/docs.md`.

2. Restart OpenCode

The plugin will discover and register your skill.

3. Use the Skill

skills_my_skill

The Agent receives the skill content and follows its instructions.

Skill Structure

Required: SKILL.md

Every skill must have a SKILL.md file with YAML frontmatter:

---
name: skill-name # Must match directory name
description: What this skill does and when to use it (min 20 chars)
license: MIT # Optional
allowed-tools: # Optional (parsed but not enforced)
  - read
  - write
metadata: # Optional key-value pairs
  version: "1.0"
---

# Skill Content

Your skill instructions in Markdown format.

Optional: Supporting Files

my-skill/
├── SKILL.md              # Required
├── scripts/              # Executable code
│   └── helper.py
├── references/           # Documentation to load as needed
│   └── api-docs.md
└── assets/               # Files used in output
    └── template.html

Skill Naming

Directory	Frontmatter Name	Tool Name
brand-guidelines/	brand-guidelines	skills_brand_guidelines
tools/analyzer/	analyzer	skills_tools_analyzer

Rules:

• Directory name: lowercase with hyphens (my-skill)
• Frontmatter name: must match directory name exactly
• Tool name: auto-generated with underscores (skills_my_skill)

Controlling Skill Access

By default, all discovered skills are available to all agents. Use OpenCode's tool configuration to control which agents can access which skills.

💡 Tip: Use skills*: false at project level to prevent context pollution, then enable only what each agent needs.

Project Level (Global Defaults)

Disable all skills by default, then enable specific ones in your opencode.json:

{
  "$schema": "https://opencode.ai/config.json",
  "tools": {
    "skills*": false,
    "skills_my_skill": true
  }
}

Per-Agent Access

Override defaults for specific built-in agents (like build, plan, etc.):

{
  "$schema": "https://opencode.ai/config.json",
  "tools": {
    "skills*": false
  },
  "agent": {
    "build": {
      "tools": {
        "skills_document_skills_docx": true,
        "skills_document_skills_xlsx": true
      }
    }
  }
}

Now only the build agent has access to document skills.

Subagent Access

For custom subagents, control tools via YAML frontmatter in the agent definition:

mode: subagent
description: Content creator agent
tools:
  skills_brand_guidelines: true
  skills_writing_style: true

This subagent gets specific skills even if they're disabled globally.

How It Works

The plugin uses Anthropic's message insertion pattern to deliver skill content:

1. Skill loading message - Announces skill activation
2. Skill content message - Delivers instructions with base directory context
3. Tool confirmation - Returns "Launching skill: {name}"

Both messages use noReply: true, so they appear as user messages (not tool responses). This ensures skill content persists throughout long conversations, even when OpenCode purges tool responses to manage context.

Path Resolution

Skills can reference files with relative paths:

Read `references/api.md` and run `scripts/deploy.sh`

The Agent receives base directory context:

Base directory for this skill: /path/to/.opencode/skills/my-skill/

And automatically resolves paths like: /path/to/.opencode/skills/my-skill/references/api.md

Troubleshooting

Skills not discovered?

• Verify SKILL.md files exist in discovery paths
• Check console for discovery messages
• Confirm frontmatter is valid YAML

Tool not appearing?

• Ensure name field matches directory name exactly
• Check for duplicate tool names (logged as warnings)
• Restart OpenCode after adding/modifying skills

Paths not resolving?

• Check the base directory shown in skill output
• Verify supporting files exist at specified paths
• Ensure paths in SKILL.md are relative (not absolute)

Invalid skill errors?

• Name must be lowercase with hyphens only ([a-z0-9-]+)
• Description must be at least 20 characters
• Name in frontmatter must match directory name

Plugin not updating?

• Check version: cat ~/.cache/opencode/node_modules/opencode-skills/package.json | grep version
• Force update: rm -rf ~/.cache/opencode then restart
• Pin version: Add @version to plugin name in opencode.json

API Reference

The plugin exports a single function that registers skills as dynamic tools:

export const SkillsPlugin: Plugin

Discovery: Scans .opencode/skills/, ~/.opencode/skills/, and ~/.config/opencode/skills/
Validation: Enforces Anthropic Skills Specification v1.0
Tool naming: skills_{name} with underscores for nested paths

See types for full interface definitions.

Advanced

Design Decisions

Why Agent-Level Permissions?

The allowed-tools field in skill frontmatter is parsed for Anthropic spec compliance, but enforcement happens at the OpenCode agent level (see Controlling Skill Access). This provides:

• Clearer permission model aligned with OpenCode's existing system
• Centralized control in opencode.json rather than scattered across skills
• Flexibility to override permissions per-project or per-agent

No Hot Reload

Skills are discovered at startup and cached. Adding or modifying skills requires restarting OpenCode. This is acceptable because skills change infrequently and simplifies the implementation.

Contributing

Contributions welcome! Fork, create a feature branch, and submit a PR.

License

MIT - see LICENSE

References

• Anthropic Skills Specification
• OpenCode Documentation

---

Not affiliated with OpenAI or Anthropic. This is an independent open-source project.