MCP Servers

MCP stands for Model Context Protocol. In plain terms: it lets the AI Assistant talk to outside tools — like a documentation search engine, your GitHub account, or a folder on your computer — so it can give you answers grounded in your actual context instead of relying on training data alone.

When you connect an MCP server, its tools become available to the AI automatically. Ask "search my repo for auth-related code" and the AI will reach through the GitHub MCP server, not guess.

The MCP badge in the AI Terminal toolbar shows how many tools are currently active (e.g. "5 tools"). Click it to open the MCP Manager.

What comes built-in

CtrlOps ships with three default servers. They appear in the MCP Manager automatically, but they start turned off until you choose to enable them.

If you delete a default server, it won't come back. You can always add it again manually.

Connect your first MCP server

The fastest path is enabling one of the defaults. Here's Context7 as an example — the steps are the same shape for the others.

How the AI uses MCP tools

You don't tell the AI to use MCP. You just ask your question naturally, and the AI decides whether a connected server can help.

What you see

1. You ask: "What changed in the auth middleware since last week?"
2. The AI emits an MCP request — you'll see a small inline panel appear in the chat showing which server is being used and what it's doing.
3. The Subagent runs — CtrlOps spins up an isolated AI loop (called a Subagent) that figures out the exact tool names and arguments, executes them, and collects results.
4. Result appears — the final answer lands back in your chat, citing what it found.

The Subagent panel shows each "turn" live: thinking step → tool call → result. It caps at 10 turns. If something goes wrong, you see the error right there instead of a silent failure.

Manage servers and tools

The MCP Manager drawer

Click the MCP badge anytime to reopen the drawer. Each server card shows:

• Status dot — green (connected), amber (connecting), red (error)
• Tool count — total tools vs. active tools (e.g. "8 tools (6 active)")
• Toggle — turn the whole server on or off
• Test — ping the server and refresh the tool list
• Edit — change name, description, URL, headers, or command
• Delete — remove permanently (asks for confirmation)

If a server errors, its toggle turns into a RETRY button in red. Click it to attempt reconnection.

Enable or disable individual tools

Click the "X tools (Y active)" line on a connected card to expand the tool list. Each tool has its own switch. Disabled tools won't be offered to the AI, which is useful when a server has many tools and you only need a few.

Add a custom server

Beyond the defaults, you can connect any MCP-compatible server.

{
  "mcpServers": {
    "my-docs": {
      "type": "http",
      "url": "https://docs-api.example.com/mcp",
      "description": "Internal company documentation",
      "headers": {
        "Authorization": "Bearer ${input:api_key}"
      }
    }
  },
  "inputs": [
    { "id": "api_key", "description": "API Key", "password": true }
  ]
}

Security

Validation guards in place:

• URLs must start with http:// or https:// and have a valid domain with a TLD
• HTTP URLs that aren't localhost trigger a security warning
• Stdio commands are checked against a whitelist of known executables (npx, node, python, docker, uvx, etc.)
• Secret formats are enforced by regex — e.g., Context7 keys must start with ctx7sk-, GitHub tokens with ghp_ or github_pat_
• Whitespace-only submissions are rejected everywhere
• All connections time out after 120 seconds
• Server management is disabled while the AI is running to prevent conflicts

Troubleshooting

Go deeper

Tips