Claude Code Extension

Overview

The ManageLM Claude Code extension connects Claude to your infrastructure through the MCP (Model Context Protocol). Once connected, Claude can manage your servers using natural language: no scripts, no SSH, no context switching.

You describe what you need in plain English. Claude selects the right skill, sends the instruction to the agent running on your server, and returns the result. All within the conversation.

Features

Architecture

Claude communicates with the ManageLM portal via MCP. The portal dispatches tasks to agents over WebSocket. Agents execute tasks using a local LLM and report results back. No SSH, no inbound ports on your servers.

Installation

From the marketplace

claude plugin marketplace add managelm/claude-extension
claude plugin install managelm@managelm

Manual MCP configuration

Add the server from the terminal with claude mcp add --transport http managelm https://app.managelm.com/mcp, or in a project .mcp.json:

{
  "mcpServers": {
    "managelm": {
      "type": "http",
      "url": "https://app.managelm.com/mcp"
    }
  }
}

Local testing

claude --plugin-dir ./path/to/managelm-claude

Authentication

ManageLM uses OAuth 2.0 with PKCE. On first connection, Claude opens your browser to authorize access.

1. Create a ManageLM account if you don't have one

2. Go to Settings > MCP Credentials in the portal and enable them

3. When Claude connects for the first time, a browser window opens for authorization

4. Approve access: tokens are managed automatically from here

Access tokens expire after 30 minutes and are refreshed automatically. You should never need to re-authenticate manually.

Skill-Based Tools

Each skill accepts a target (hostname, group, or "all") and a free-text instruction describing what you want done. Claude selects the appropriate skill automatically based on your request.

Core skills

Extended skills

Additional skills can be imported from the built-in catalog using the available_skills tool:

Discovery & Management Tools

These 32 built-in tools let Claude explore your infrastructure, run scans, check task status and manage operations. Some only appear when they apply: the monitor, certificate, backup, credential and keystore searches follow the optional features enabled on your account, and the hosting tools appear once a hosting connector exists.

Targeting

Every skill-based tool accepts a target parameter that determines which server(s) to run the task on:

Claude automatically resolves targets. If you say "Check disk usage on all production servers", Claude uses the group name production as the target.

Slash Skills

The extension includes two user-invocable skills you can trigger directly:

/managelm:setup

Walk through connecting Claude to your ManageLM portal. Guides you through configuring the MCP endpoint, authenticating, and verifying the connection.

/managelm:troubleshoot

Diagnose connection issues, task failures, and agent problems. Works through connection checks, agent status, task errors, and tool availability in order.

Example Usage

Just describe what you need in natural language. Claude handles skill selection, targeting, and execution:

# System info
> Check disk usage on web-prod-01
> Which servers have CPU usage above 80%?
> Show me the last 50 lines of /var/log/syslog on monitoring-1

# Package management
> Install nginx on all servers in the staging group
> What packages are installed on web-prod-01 that match "python"?
> Update all packages on db-primary

# File operations
> Upload my local config.yml to /etc/myapp/config.yml on db-primary
> Show me the contents of /etc/nginx/nginx.conf on lb-01

# Security
> Run a security audit on all production servers
> Search for any security findings related to SSH across all servers
> Harden SSH configuration on web-prod-01

# Services
> Restart nginx on lb-01
> Show me all failed systemd units on db-primary

# Infrastructure discovery
> List all my servers and their status
> Which servers are running Docker?
> Search for any server with PostgreSQL installed

Security

• OAuth 2.0 with PKCE: secure browser-based authorization, no client secrets stored locally
• Short-lived tokens: access tokens expire after 30 minutes, automatic refresh
• Command allowlisting, each skill defines exactly which commands are permitted. The agent enforces this in code, not just prompts.
• Local LLM execution, agents use a local LLM (on the server itself or a dedicated host). Your data never leaves your infrastructure.
• Zero inbound ports, agents connect outbound to the portal via WebSocket. No SSH, no open ports on your servers.
• Ed25519 signatures, every task dispatched to an agent is cryptographically signed
• Read-only by default, the base skill has no write commands. Mutating skills must be explicitly assigned by an admin.
• Full audit trail: every task is logged with user, agent, skill, instruction, result, and any file changes
• Task revert, file changes are tracked via git and can be reverted at any time

Requirements

• Claude Code 2.0.12+ for the plugin, or any MCP client with remote HTTP servers (Claude Desktop, claude.ai)
• ManageLM account, create one (free for up to 10 agents)
• ManageLM Agent, installed on each Linux server you want to manage

Troubleshooting

401 Unauthorized

The sign-in did not complete, or the token was revoked (for example after you reset your MCP credentials in Settings > MCP & API). Run /mcp, select the ManageLM server (listed as plugin:managelm:managelm when it comes from the plugin) and authenticate again.

Connection refused or timeout

Verify the portal URL in your MCP configuration is correct and reachable. The endpoint should be https://your-portal.com/mcp.

No agents listed

Either no agents are enrolled, or your account role restricts agent visibility. Check the portal dashboard to verify agents are enrolled and your user has access.

OAuth flow not starting

Ensure the MCP server is configured with "type": "http" (not "stdio"). Plugins need Claude Code 2.0.12 or later.

Skill not available on agent

Use get_agent_skills to see what's assigned. Skills must be assigned to agents by an admin in the portal. Use list_available_skills to discover importable skills from the catalog.

Task timeout

Regular tasks have a default timeout (configurable in portal settings). File transfers have a 300-second timeout. For large operations, break the task into smaller steps.

Tools not appearing in Claude

The MCP connection may not be established. Check the MCP status in Claude Code. Try /managelm:troubleshoot for guided diagnostics, or run /reload-plugins to refresh.