io.github.divitkashyap/auto-skill-loader

Official

by divitkashyap · Python

Auto-loads skills into agent sessions. MiniMax vision and search proxy.

Dynamically loads skills into agents with vision and search capabilities.

README.md9 KBView on GitHub

auto-skill-loader

Give your AI agent a persistent skill library it auto-loads at session start.

OpenCode image note: OpenCode doesn't yet expose real file paths when you paste images — the agent sees a filename instead. Workaround: give the agent an actual file path instead of pasting. Skills auto-load at session start as expected.

mcp-name: io.github.divitkashyap/auto-skill-loader

How it works

You configure which skills to auto-load in ~/.config/auto-skill-loader/config.yaml
The MCP server reads skill files from your skills directory and exposes them via skills://active
At session start your agent reads skills://active and gets all approved skills auto-injected
No explicit triggers needed — the agent already knows your skills

Why

The main problem: In OpenCode, skills don't auto-load — the agent has to explicitly invoke them or rely on unreliable fuzzy matching. auto-skill-loader solves this by exposing your skill library as an MCP resource that gets read at session start. Deterministic, no guessing.

Bonus: Our proxy tools also work around OpenCode's broken MCP stdio transport for the MiniMax vision and web search tools.

Installation

Option 1: uvx (recommended — no install needed)

bash

uvx auto-skill-loader

Option 2: pip

bash

pip install auto-skill-loader
auto-skill-loader

Option 3: Build from source

bash

git clone https://github.com/divitkashyap/auto-skill-loader.git
cd auto-skill-loader
pip install -e .
auto-skill-loader

Configuration

OpenCode

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

json

{
  "mcp": {
    "auto-skill-loader": {
      "type": "local",
      "command": ["uvx", "auto-skill-loader"],
      "enabled": true
    }
  }
}

Claude Code

Note: Claude Code requires a specific JSON format via add-json:

bash

claude mcp add-json -s user auto-skill-loader '{"type":"stdio","command":"/FULL/PATH/TO/python","args":["-m","server"],"env":{"MINIMAX_TOKEN_PLAN_KEY":"sk-cp-YOUR-KEY-HERE"}}'

Replace /FULL/PATH/TO/python with the path to your Python (e.g. /Users/YOU/auto-skill-loader/.venv/bin/python).

Or for uvx (requires network on first run):

bash

claude mcp add -s user --transport stdio -e MINIMAX_TOKEN_PLAN_KEY=sk-cp-YOUR-KEY auto-skill-loader -- uvx auto-skill-loader

Tested On

Host	Skill Auto-Load	Vision Proxy	Image Paste	Notes
Claude Code (macOS)	✅	❌ Not needed	✅	Use `minimax-token-plan` MCP directly instead
OpenCode (macOS)	✅	✅ Works	❌ Broken	Give file paths instead of pasting

Skill auto-loading confirmed working on both hosts. The vision proxy bypasses OpenCode's broken stdio transport, but image paste-to-path is a separate rendering issue in OpenCode that requires using file paths.

Other MCP-compatible hosts (Cursor, Zed, etc.) should work with the same configuration — contributions welcome.

Platform Differences & Known Issues

Image Input: OpenCode vs Claude Code

Both hosts work with auto-skill-loader vision tools, but image input behaves differently:

Host	How images are passed	Recommended workflow
Claude Code	Images uploaded to URL automatically → tool receives URL	Paste image directly ✅ works
OpenCode	Inline images render visually but may not give tools a real path	Give a file path instead of pasting

OpenCode note: When you paste an image in OpenCode, it may render inline but the agent sees it as a filename string (e.g. logo.png) rather than a real filesystem path. This is a known OpenCode rendering behavior.

Workaround for OpenCode: Instead of pasting, give the agent the actual file path:

code

analyze this image: /path/to/your/image.png

The agent can access local files directly in OpenCode. If the image is only in your clipboard, the agent can extract it to /tmp/ first.

What We're Monitoring

We actively track the following OpenCode issues:

Inline image rendering (images pasted don't expose real paths to tools)
MCP stdio transport for local servers (our proxy tools work around this)
Session persistence of skills across restarts

If OpenCode releases a fix for inline image paths, this documentation and the vision-analysis skill will be updated to support direct paste workflow. We monitor OpenCode's changelog and release notes — check this repo for updates.

Other Known Issues

Issue	Severity	Workaround
OpenCode inline images show as filename, not path	Medium — affects paste workflow	Use file paths instead
Claude Code auth conflict (ANTHROPIC_AUTH_TOKEN vs managed key)	Low — cosmetic warning	Harmless, can be ignored
First vision call may take 3-5s (uvx download)	Low — one-time	Subsequent calls are ~200ms

Setup

Create skills directory (symlink to your existing skills):

bash

mkdir -p ~/.config/auto-skill-loader
ln -sf ~/.config/opencode/skills ~/.config/auto-skill-loader/skills

Edit ~/.config/auto-skill-loader/config.yaml:

yaml

active_skills:
  - vision-analysis
  - context-maintainer
  - markdown-mcp
skills_dir: ~/.config/auto-skill-loader/skills

Restart your agent. It will now auto-load all listed skills at session start.

Tools

Tool	What it does
`list_skills`	List all available skills in skills_dir with descriptions
`activate_skill`	Add a skill to your approved list (persists to config.yaml)
`deactivate_skill`	Remove a skill from your approved list
`get_skill_info`	Get details about a specific skill
`get_active_skills`	List currently active skill names
`suggest_skills`	If no skills are active, suggests common ones to get started
`check_prerequisites`	Validate a skill's dependencies (MCP tools, API keys, env vars)

Bonus: MiniMax Vision & Web Search Proxy

auto-skill-loader also exposes two tools that proxy to minimax-coding-plan-mcp with a working stdio transport:

Tool	What it does
`minimax_understand_image`	Analyze images (JPEG, PNG, GIF, WebP up to 20MB)
`minimax_web_search`	Web search using MiniMax

Note: For vision specifically, mmx-cli is now the recommended approach — it's a direct REST call to MiniMax VLM, no MCP transport issues, and handles URLs natively. These proxy tools are useful if you want a unified MCP interface for both skill loading and vision.

The OpenCode MCP Bug

When OpenCode's built-in minimax-coding-plan-mcp MCP integration (minimax-token-plan) is configured, the understand_image tool fails with:

code

API Error: login fail: Please carry the API secret key in the 'Authorization' field

This happens even when:

✅ MINIMAX_API_KEY / MINIMAX_TOKEN_PLAN_KEY is set correctly
✅ API key is valid (same key works via direct API calls)
✅ Token Plan has available vision quota

Root cause: OpenCode's stdio transport for local MCP servers sends messages in a way that breaks the MCP protocol — likely batched writes without proper flush between JSON-RPC messages. Direct subprocess tests with sequential writes + flush() work fine.

The fix: Our proxy tools in auto-skill-loader use proper sequential stdio communication, bypassing OpenCode's broken transport layer.

Setup

Set your MiniMax Token Plan key in ~/.config/opencode/.env:

bash

MINIMAX_TOKEN_PLAN_KEY=sk-cp-your-key-here

Add auto-skill-loader to ~/.config/opencode/opencode.json:

json

{
  "mcp": {
    "auto-skill-loader": {
      "type": "local",
      "command": ["/path/to/venv/bin/python", "-m", "server"],
      "enabled": true
    }
  }
}

Critical: If you have minimax-coding-plan-mcp configured directly in opencode.json (the minimax-token-plan entry), remove or disable it — its broken stdio transport will cause "login fail" errors. The proxy tools in auto-skill-loader replace it entirely.
Restart OpenCode and verify: /ask Do you have auto-skill-loader_minimax_understand_image available?

Diagnosis

If you see "login fail" errors after setup:

Disable the broken minimax MCP — ensure "minimax-token-plan": { "enabled": false } or remove it entirely
Restart OpenCode completely — MCP servers are re-spawned on each session
Check with: /ask Call minimax_understand_image with image_source="/any/real/image.png" and prompt="test"

Resources

Resource	What it does
`skills://active`	All approved skill contents concatenated — read by host at session init
`skills://config`	Your current config.yaml content

Security

User-controlled — only skills in config.yaml are loaded
No network fetches — everything is local
No prompt injection — skills come from your own configured directory

Repo Structure

code

auto-skill-loader/
├── src/
│   └── server.py        # MCP server (Python stdlib + mcp package)
├── pyproject.toml       # Package config
├── README.md            # This file
├── SKILL.md             # For agent onboarding
└── LICENSE              # MIT

Requirements

Python 3.9+
mcp package (pip install mcp)
pyyaml package (pip install pyyaml)

Or just use uvx auto-skill-loader which fetches dependencies automatically.

Install

Configuration

Environment variables

MINIMAX_TOKEN_PLAN_KEYrequiredsecret

MiniMax Token Plan API key from https://platform.minimax.io

claude_desktop_config.json

{
  "mcpServers": {
    "auto-skill-loader": {
      "command": "uvx",
      "args": [
        "auto-skill-loader"
      ],
      "env": {
        "MINIMAX_TOKEN_PLAN_KEY": "<YOUR_MINIMAX_TOKEN_PLAN_KEY>"
      }
    }
  }
}