Foundations·15 Oct 2026·6 min read

Your AI Assistant Needs Server Access. Don't Hand It the SSH Key.

The moment an AI assistant needs to touch a server, most teams face two answers: say no, or paste a credential into its config. Both are bad. There is a third: the assistant connects to your PAM, works as the person who approved it, and every session is recorded under both names.

Your AI Assistant Needs Server Access. Don't Hand It the SSH Key.

Your engineers' AI assistants are already on their laptops. They read the codebase, they write the pull requests, and sooner or later one of them hits a task that ends with: "now check the disk on that server." At that moment most organizations discover they have exactly two answers, and that both are bad.

The first answer is no. The assistant stays locked out of anything real, the engineer copies terminal output into the chat window by hand, and the productivity the company paid for stops at the first firewall. Bans have a short shelf life; what actually happens next is answer number two.

The second answer is handing over a credential. An SSH key lands in the assistant's config file, or a database password goes into an environment variable it can read. Now a piece of software that writes its own commands holds a standing secret, outside the vault, outside rotation, invisible to the audit trail. Every control your PAM enforces on people has just been bypassed by the newest and least predictable actor on the network.

The third answer

The structural fix is the same one PAM applied to humans years ago: nobody carries credentials to the target, everybody goes through the broker. Monopam now extends that to AI assistants over MCP, the open protocol that tools like Claude Code, Cursor, VS Code, and Claude Desktop already speak.

The setup is one line: the assistant gets the URL of your Monopam MCP server. The protocol handles the rest. The assistant discovers who guards the resource, registers itself, and opens a browser window where the engineer signs in with their corporate account: standard OAuth with PKCE, through Monosign. Then a consent screen asks the one question that matters: this assistant will act for you; allow?

Note what never happened in that flow. No password was typed into an assistant. No key was pasted into a config file. The assistant ends up holding a token, and that token is the interesting part.

An extension of the person, not a new user

The token carries the engineer's identity, with the assistant's identity layered on top. Every policy decision applies to the person: the assistant can open exactly the sessions its owner could open, and nothing more. There is no new "service account" with its own quietly accumulating permissions; there is a person, and a tool acting visibly on that person's behalf.

From there, the assistant works through the same brokered path as everyone else: SSH, database, RDP, VNC, and Kubernetes sessions, access requests, secrets. Credentials come out of the vault and go to the target; the assistant never sees them. A secret read is separately audited as a reveal.

And everything is on the record. Sessions opened by an assistant appear in the live list with an AI badge and are recorded like any other privileged session. The audit trail names both parties: Claude Code (monopam) on behalf of the engineer. When the auditor asks how AI touches production, the answer is the same screen you already show them.

Approvals, and the off switch

Autonomy needs limits, and the useful limits are the ones that already exist. A command that matches a privileged-command rule does not run; it goes to the approval flow, and the approval is bound to that exact command text and spent once. Connection-level approval works the same way.

Control cuts in the other direction too. The engineer can disconnect their own assistant at any time from their account. An administrator can suspend it centrally, and here is the detail worth measuring: access stops within two minutes, without waiting for a token to expire. We timed it at 91 seconds from suspend to the assistant's next call being refused.

Why this beats both old answers

Compared to "no": the engineer keeps the assistant, and the work the assistant does on servers is faster to review than the work an unassisted human does, because every step is recorded and attributed.

Compared to the SSH key: there is no standing credential to steal, no shadow account to forget, and no gap in the audit trail. The blast radius of a compromised assistant is the permissions of one person, bounded by approval rules, with a kill switch on top.

Banning AI is not a strategy. Neither is opening everything up. The third option is treating the assistant like what it is: a powerful tool acting for a person, and governing it exactly as tightly as you govern the person.

Setup takes minutes: see the walkthrough for connecting Claude Code, Cursor, or VS Code to Monopam, or start with the MCP Server overview.

Tagsmcpai-agentsprivileged-accesspammonopam

Ready to start managing
identities the right way?

Spin up a fully-loaded trial tenant in under five minutes. No credit card. No sales gate.