Share secrets with AI agents

Never paste a key into the chat.

The agent asks, you answer on a seal.net page, and it gets a file path.

npx -y sealnet-mcp@latest

MCP server · Claude, Cursor, Cline and any MCP host

The key pasted into a chat.

If you’ve had to give an AI agent an API key recently, you probably pasted it into the chat window. The key stays there in the transcript and in some probably less visible places in the back-end logs. Anything inside the context of an AI model can surface later in a log or a tool call or even as part of a random answer.

You can do better, and it isn’t hard. SEAL lets agents ask for keys and people give them in a way that never involves anything appearing in the chat. Nothing in between ever holds the secret.

Asked, not pasted

A seal.net page, never the chat.

A path, not the value

The answer lands in a file only its owner can read.

Straight into a program

seal run hands the key to one program, wipes the file and keeps the value out of the output.

How an agent asks.

  1. Add the server

    npx -y sealnet-mcp in Claude, Cursor or any MCP host. No setup.

  2. The agent asks

    It calls seal_request with what it needs and where that key is created.

  3. You answer on the page

    A seal.net page: on your computer it opens in your browser; elsewhere you get its link, or a card in your inbox.

  4. It gets a path

    The value goes to a private file, or into one program with seal run.

How the agent asks.

When an agent needs a key, it asks for it. The default tool is called seal request. If the agent is running with the SEAL MCP server enabled, it just calls the request MCP tool.

The agent can name the secret and where to find it, and that becomes the seal.net page that asks the person for the secret. The page also shows the address of the page where the secret is created, so you can find it in one click, and the binding of a key: the hosts it may go to and for how long.

Where the value goes.

When the answer is entered, it’s written to a file owned only by you, and the agent is given the path to that file. The agent is never told the actual value of the secret.

When the agent wants to use the secret, it does so by running the program that needs it via the command seal run. The secret is written to the environment of the program being run, and immediately afterward the file is deleted. When the program prints the value, seal run replaces it with a placeholder before the output reaches the agent, and an agent with the MCP server gets the same from seal_run.

When the key is somewhere else.

When the agent runs on your computer, the page opens in your browser by itself; when it runs in the cloud, you get its link, or a card in your inbox. The eight words on the seal.net page match the ones the agent shows in the chat. That lets you verify that you didn’t get a phishing link.

If the agent wants to ask a colleague of yours for a secret, you have to forward the request link to your colleague. Any one who is given the link can see the request, so you shouldn’t ask a random stranger. The person who receives it sees only the request, and the answer is encrypted to the public key of the agent, so even the SEAL server can’t read the answer.

If the agent is connected to your SEAL inbox, the request can wait there as a card: you open box.seal.net and answer it on the same seal.net page, with no link to pass around.

The honest limit.

The whole point of the seal request command and the seal.net pages is to get the secrets out of the chat, out of both the chat window and the logs of the chat server. SEAL keeps secrets out of all those places, but it doesn’t keep them away from the agent that is asking.

The seal.net server can’t read the answer it returns to an agent, but the agent can do what it wants with the secret it’s given, including storing it. So you shouldn’t give secrets to agents you don’t trust.

Questions, answered.

Does the model see my key?
SEAL gives the agent a file path, not the value, and never prints the value. What happens next depends on the agent: seal run passes the key to a program without showing it.
What if the agent runs in the cloud?
It sends you a seal.net page with the same request, or puts it in your SEAL inbox if you connected one. Check that the eight words on the page match the ones in the chat, then answer there.
Can a colleague provide the key?
Yes. Choose to ask another person: they get a link, and their answer is encrypted to the agent’s key.
Does SEAL protect the key from the agent?
No. It keeps the key out of the chat, the logs and the SEAL server. The agent you gave it to can store or use it, so give keys only to agents you trust with them.
npx -y sealnet-mcp@latest

MCP server · Claude, Cursor, Cline and any MCP host

Read the documentation

One seal. Every platform.