Agent Plugins

An Agent Plugin is a folder on the LibreChat server that bundles Skills, MCP servers, and optional command hooks that belong together. An operator drops the folder into the plugin directory and restarts LibreChat; there is nothing to install from the UI. LibreChat follows the open Agent Plugins 1.0.0 package format, so one package can be shared with other tools that support it.

Agent Plugins are experimental. The package format and behavior may change.

What users see

Users never see a plugin as such. They see what it contains:

• Its Skills appear in the Skills list next to other skills. They are read-only, follow the usual Skills permissions, and can be switched on or off with Available to agent like any other skill.
• Its MCP servers appear in the Tool Library and can be added to agents like servers defined in librechat.yaml.
• Its hooks run on the server at points in an agent run, such as before a tool call or when the run stops. A hook can block a tool call, add context to the prompt, or stop a run, and the user only sees the effect.

If a skill or MCP server name collides with one defined elsewhere, the other definition wins: a skill in the standalone deployment skills directory beats a plugin skill, and a server in librechat.yaml beats a plugin server.

What a plugin can contain

A plugin folder has a plugin.json manifest and any of:

• skills/, one folder per skill, each with a SKILL.md.
• mcp.json, MCP server definitions. Local (stdio) servers run from inside the plugin folder and get their own persistent data directory.
• ai.librechat/hooks/hooks.json, command hooks. These run as processes on the API host and are disabled unless the operator turns them on.

Each part loads on its own, so a broken mcp.json does not stop the plugin's skills from loading. Problems are reported in the startup logs.

Setup

Plugins load from ./plugin at the project root by default, and command hooks are off by default. See Agent Plugins configuration for the directory settings, the package layout, manifest and MCP validation rules, and the hook execution contract.