Security and privacy

Real access for your AI. Real limits, too.

SolutionMCP gives AI apps real access to your computer, so it's built around three ideas: ask before risky actions, limit what can be reached, and keep a record. Here's how, and where its limits are.

Ask before risky actions

Deleting, overwriting and running commands need your OK in the Balanced mode. The question shows what will run, and no answer is a no.

Limit what can be reached

File tools work only in the folders you allow, and keys, passwords and private data stay blocked even there.

Keep a record

Every tool call is written to a log on your computer, with secrets masked, so you can always see what was done.

What leaves your computer

SolutionMCP itself sends nothing anywhere, except:

  • What your tools do, like reading a web page you asked about.
  • The results it gives your AI app, which the app usually sends on to its AI service, like anything you type.
  • The update check, which reads the list of releases from solution-mcp.com.
  • Reachable from anywhere, if you turn it on: your computer stays connected to SolutionMCP's relays, which pass encrypted connections along and count how much data goes through.
  • License checks: using a license on a computer, a check for renewals about once a week, and how the license stands. These send the license key or ID, the computer's ID and name, and the version, and nothing about your tools, files or what your AI apps did. The license itself is checked on your computer, offline.

There's no analytics, no telemetry and no account. So the rule of thumb is simple: don't have the AI read a file into a conversation you wouldn't paste into it yourself.

Files: folders, real paths and blocked places

  • Allowed folders. File tools only reach the folders you allow (your home folder to start).
  • Real paths. Paths are followed to where they really point before they're checked, so a link inside an allowed folder can't reach outside it.
  • Blocked even inside allowed folders: SSH and GPG keys, cloud credentials (AWS, Azure, Google Cloud, Kubernetes, Docker), .netrc, git and package-manager credentials, password stores, .env files, private keys, and on each system its keychains, credential stores, and browser passwords and cookies.
  • Deleting goes to the Trash or Recycle Bin, so mistakes can be undone.

SolutionMCP protects itself

The AI can't loosen its own permissions. File tools can't change SolutionMCP's settings, packs or log, and can't even read its secrets, keys or tokens. The dashboard needs a one-time sign-in link the AI can't get. The one thing the AI can add is a new tool, and that always shows you the whole tool and asks first.

The terminal

The shell tool runs real commands, so in the Balanced mode it asks before every one. Commands you allow ahead of time must be simple ones: anything chained with ;, |, & or redirects always asks. Dangerous commands like sudo, shutdown and disk-wiping ones are always refused. Your own tools are safer still: values from the AI are passed separately or quoted for their language, so they can't inject commands.

Apps that connect over the web

ChatGPT, claude.ai and other apps that connect by URL sign in with OAuth. You approve each app in your browser with a one-time code only you can see, and the page shows who's asking. Tokens are short-lived and tied to the address they were issued for, and you can disconnect an app at once.

Your computers, paired

  • Encrypted, and bound to the right computer. Each computer has its own certificate and proves who it is on every connection.
  • The pairing code can't be guessed offline. It never crosses the network, works once, for 10 minutes, and stops after 5 wrong tries.
  • Only that computer's own tools. Not its dashboard, and not the tools it uses from other computers.
  • Reachable from anywhere goes through SolutionMCP's relay, which only passes encrypted traffic along by name: the encryption is all the way to your computer, which holds its own certificate, so the relay can't read or change it. Only apps you approve with the code your computer shows get in.

Updates you can trust

SolutionMCP checks solution-mcp.com for updates once a day, if you let it. The list of releases is signed with SolutionMCP's own release key, built into every copy, and a list that isn't signed with it is refused. Each download must match the checksum in that signed list before it's installed. On a Mac, the app is signed and notarized by Apple.

What it can't protect against

We'd rather you know:

  • Tricks hidden in content. A web page, email or document can carry instructions aimed at the AI. Keep approvals on for anything that deletes or overwrites.
  • Approving the wrong thing. Read approval questions: they show what will run.
  • AI agents with their own terminal. Coding agents that run commands themselves act as you, outside SolutionMCP. Keep their own approvals on.
  • Trusted mode and broad "Always run" rules turn the questions off. Use them only where you're comfortable with that.

Report a security issue

If you find a security problem in SolutionMCP or this website, please email security@solution-mcp.com with the details and how to reproduce it. Please give us a chance to fix it before telling others. Fixes for serious security problems reach every license, even after its year of updates.

The technical details are in the manual: Privacy and security, Approvals, Folders the AI can use and Permission modes.