Use skills from your IDE
Every skill in the library installs with one command from its page. This page is about the step before that: your agent asking aiqbench which skills fit the task you just typed, and offering them to you.
What happens
- 1
You type a task
“Extract the tables from these PDF invoices.” With the plugin, every prompt is checked; with the MCP server, the agent asks when a task looks like it needs a tool.
- 2
aiqbench suggests
At most three skills, each with its trust score, one line on why it helps this task, and what it costs in context. Nothing is installed or used yet.
- 3
You choose
Use a skill once for this task, install it for next time, or say no. If you say nothing, the agent carries on and does not mention skills again for that task.
Only skills that passed the scanners with a trust score of 70 or more are ever suggested; community skills also need a specific trigger, not a description that matches everything.
Add the MCP server to your agent
One endpoint for every agent that speaks the Model Context Protocol. Pick yours; the snippet is ready to paste.
Claude Code
terminalclaude mcp add --transport http aiqbench https://aiqbench.com/mcp
Cursor
~/.cursor/mcp.json{
"mcpServers": {
"aiqbench": { "url": "https://aiqbench.com/mcp" }
}
}Codex
~/.codex/config.toml[mcp_servers.aiqbench] url = "https://aiqbench.com/mcp"
VS Code · GitHub Copilot
.vscode/mcp.json{
"servers": {
"aiqbench": { "type": "http", "url": "https://aiqbench.com/mcp" }
}
}Gemini CLI
~/.gemini/settings.json{
"mcpServers": {
"aiqbench": { "httpUrl": "https://aiqbench.com/mcp" }
}
}JetBrains AI Assistant · Windsurf
Settings → MCP → add server{
"mcpServers": {
"aiqbench": { "url": "https://aiqbench.com/mcp" }
}
}The server exposes three tools: find_skills (suggestions for a task), get_skill (read a skill to follow it once, without installing) and install_command (the one-line install for your agent).
Claude Code: the plugin
The plugin adds the MCP server, the find-skills skill, and a hook that checks every prompt of four words or more. Suggestions arrive within 1.5 seconds and never hold up your prompt.
/plugin marketplace add ipbardasu/my-news /plugin install aiqbench@aiqbench
Two commands inside Claude Code. The marketplace lives in the aiqbench repository; if the install step says it cannot read it, add the MCP server above instead — it is the same server.
Any agent: the find-skills skill
A small skill that tells your agent when it is worth asking the library (a file format, a tool, an audit, a multi-step job) and how to offer what comes back. It uses the MCP server when you have added it, and the public API otherwise.
npx skills add https://aiqbench.com/s/find-skills -a claude-code
Change the agent at the end of the command for Cursor, Codex, GitHub Copilot or Gemini CLI. Read the skill.
Settings
Set these in the shell your agent runs in. They apply to the plugin's hook; the MCP tools take the same choices as parameters.
| Setting | Effect |
|---|---|
| AIQBENCH_SUGGEST=on | The default: the text of your prompt is sent to aiqbench to find matches. |
| AIQBENCH_SUGGEST=keywords | Only the nouns, identifiers and file types of your prompt are sent, never the sentence. Good for confidential work; matches are a little less precise. |
| AIQBENCH_SUGGEST=off | The hook sends nothing. The MCP server and the skill stay available when you ask. |
| AIQBENCH_NEVER=pdf,xlsx | Skills never to suggest. |
| AIQBENCH_URL=https://… | Another aiqbench instance, for example one you run yourself. |
What leaves your machine
The task text (or its keywords) is sent to aiqbench to find matches and is not stored: the only trace is a hashed cache key kept for a day, alongside the one-line explanation written for a suggestion. No account is needed and none is linked to these requests.