AI Agents & MCP

Let an AI agent operate your machines directly. Folderbird runs a hosted Model Context Protocol (MCP) server, so Claude, OpenClaw, or any MCP‑capable client can list your devices, run commands, and read or write files — using the same scoped, audited API keys.

What is MCP? The Model Context Protocol is an open standard for giving AI agents tools. Connect Folderbird once and the agent automatically discovers the device tools below — no plugin to build.

Connect your agent

It's a remote (hosted) MCP server — nothing to install. Point your client at the endpoint and send your API key as a bearer token:

Endpoint:  https://folderbird.com/mcp
Header:    Authorization: Bearer fbk_live_YOUR_KEY

Claude Desktop

In claude_desktop_config.json:

{
  "mcpServers": {
    "folderbird": {
      "url": "https://folderbird.com/mcp",
      "headers": { "Authorization": "Bearer fbk_live_YOUR_KEY" }
    }
  }
}

If your client only supports local (stdio) servers, bridge it: npx mcp-remote https://folderbird.com/mcp --header "Authorization: Bearer fbk_live_YOUR_KEY"

OpenClaw

In openclaw.json, then run openclaw gateway restart:

{
  "mcp": {
    "servers": {
      "folderbird": {
        "transport": "http",
        "url": "https://folderbird.com/mcp",
        "headers": { "Authorization": "Bearer fbk_live_YOUR_KEY" }
      }
    }
  }
}

Tools

Each tool needs the matching scope on the key — the agent only sees the tools its key is allowed to use.

Tool	What it does	Scope
`list_devices`	List the devices this key can reach, with online status.	devices:read
`run_command`	Run a shell command (cmd or PowerShell) and return its output.	exec
`browse_folder`	List the folders and files in a path on a device.	browse
`read_file`	Read a text file from a device (up to 256 KB).	download
`write_file`	Create a new text file, or edit an existing one by overwriting it, with the text you provide.	upload
`create_folder`	Create a folder (and any missing parent folders) on a device.	upload
`upload_file`	Upload any file (binary or text) to a device, sent as base64.	upload
`download_file`	Fetch any file — binary or large (images, PDFs, zips) — as a temporary download link.	download
`system_info`	Structured health snapshot — OS, CPU, RAM, disk, uptime, top processes — as JSON.	devices:read
`search_files`	Fast file search by name or content, via the Windows Search index.	browse

What you can ask it to do

Real prompts, grouped by what they unlock:

📥 Reach your files from anywhere

“Grab the Q3 deck from my desktop and give me a download link.”
“Find last month's invoice PDF in my Documents and send it over.”

🩺 Troubleshoot in plain English

“My PC's been sluggish — what's eating CPU, memory, and disk?”
“Read the last 50 lines of my app's error log and tell me what broke.”

🛠️ Fix & run things remotely

“The print spooler is stuck — restart it.”
“Run my nightly backup script now.”
“Clear my temp folder and tell me how much space you freed.”

📊 Check on an always‑on machine or server

“Is nginx running? Show disk space and memory.”
“What's listening on port 8080?”

📝 Quick reads & edits

“What's in my todo.txt?” → “Add ‘call the bank’ to it.”
“Read my config.json and check the database setting looks right.”

🔄 Move files between your own machines

“Copy the build output from my laptop to my desktop.”

Security

Scoped — a key that lacks exec simply won't expose run_command.
Device‑bound — bind a key to specific machines so an agent can never touch the rest.
Rate‑limited — per‑key limits protect your devices from runaway loops.
Audited — every command an agent runs is logged with the key, device, and time on your keys page.

Tip: give each agent its own key with the narrowest scopes and device set it needs. Revoke it anytime — the agent stops instantly.

Other connectors

Not using MCP? The same capabilities are available over plain HTTP for n8n (HTTP Request node), webhooks, and custom scripts — see the API Reference.