Trustwelt AI Security · Agent gateway

An agent does what it is allowed to do. Each step is decided on the computer.

Claude Code, Codex CLI, GitHub Copilot, Cursor and Gemini CLI call their hooks before and after each tool they use. On a managed computer the hook is Trustwelt's: the call reaches a small gateway next to Trustwelt Desktop, which decides in well under a millisecond and writes a record. The agent sees the decision; the person sees the questions; the organisation sees the sessions.

A module of Trustwelt AI Security, included with it. Runs on: Managed computers with Trustwelt Desktop and its browser extension: the Chromium family and Firefox on Linux, Firefox on Windows; the AI agent gateway on managed Linux computers; other programs through the network filter, sites only.

Tenant Admin › AI services › Sessions: a table of agent sessions with person, computer, routine calls, refusals and questions, and one session open in the right-hand panel with its figures and timeline: a refused connection, a refused read, a recorded connection with a credential injected, a model call sent masked.

What it does

Agent gateway, in plain words.

  • Tiers

    Five kinds of call

    Read, write, execute, destructive and exfiltration. A shell command is split into its simple commands and raised by what it does; a write outside the workspace the agent was started in is destructive.

  • Rules

    First match decides

    Rules on a tool, a command pattern or a tier: allow, allow and record, ask the person, or refuse, each with words the person reads. A question waits a set time; with no answer, the call is refused.

  • Folders

    Guardrails no rule loosens

    The agent's file tools may not read ~/.ssh, ~/.aws, ~/.config/gcloud, ~/.kube, browser profiles, the vault or other agents' credential files, and may not change its own settings files or git hooks.

  • Unknown agents

    Recorded or refused

    A program Trustwelt does not know is recorded or refused by your choice. Permission-bypass modes of the known agents are refused above a read while they are disabled.

  • Output

    Secrets masked on the way back

    What a tool returns to the agent is read for private keys, API tokens and passwords, and those are masked before the agent sees them.

  • Managed settings

    Written by Trustwelt, repaired if edited

    Each allowed agent's managed settings, hooks included, are written by Trustwelt's root helper and written again when found missing or edited; the state shows per agent in Agent analysis.

How it works

From the console to the computer.

  1. Publish the rules

    Tenant Admin › AI services › Agent rules: write the rules, test them against a sample call, then publish; every version kept, any earlier one published again; exceptions per person or group.

  2. Delivered in the signed state

    The rules reach each managed computer inside the signed web state, cut per person; the gateway reads them and compiles them once.

  3. Decided per call

    Before each tool call the hook asks the gateway; a question goes to the window of Trustwelt Desktop that started the agent, with Deny, Allow once and Allow for this session.

  4. On record

    Sessions in Tenant Admin and in the person's Portal: each decision that was not a plain allow, the questions and answers, output masked, routine calls counted; CSV and PDF.

In Tenant Admin

What your administrators see.

Tenant Admin › AI services › Sessions: a table of agent sessions with person, computer, routine calls, refusals and questions, and one session open in the right-hand panel with its figures and timeline: a refused connection, a refused read, a recorded connection with a credential injected, a model call sent masked.
A session and its timeline: decisions, questions, masked output. Never a prompt.

What is recorded, and what never is

Your organisation sees this. Nothing else.

  • DecisionsThe tool, the kind of command in two words, the tier, the rule that decided, the decision, and for a question who answered and how.
  • SessionsThe agent and its version, the workspace, start and end, counts per tool, whether the hook could reach Trustwelt Desktop.

Never:

  • A command's arguments, a file's content, a prompt or a tool's result: the record keeps the kind of command, never its text.
  • A decision taken by the browser or the agent: the gateway on the computer decides, from the signed rules.
  • A remote shell or a script run on the computer by Trustwelt.

Specification

Where it runs and how it is sold.

Runs on
Managed Linux computers with the current Trustwelt Desktop. The person's own computer: the same gateway, with the person as its administrator, from their Portal.
Agents governed
Claude Code, Codex CLI, GitHub Copilot (VS Code and CLI), Cursor, Gemini CLI; any other agent by the unknown-agents rule, the sandbox and the network rules.
Decision time
Measured on the computer: a decision in tens of microseconds, the hook's round trip in a few milliseconds.
Sold as
Part of Trustwelt AI Security.

Questions

Asked before buying.

Does this slow the agent down?
A decision takes tens of microseconds; the hook's whole round trip a few milliseconds. Routine allowed calls are counted per session rather than written one by one.
What if Trustwelt Desktop is not running?
Reads go on; anything above a read is refused with the message "Trustwelt Desktop is not running", and the miss is recorded and shown in Tenant Admin.
Can a person approve a company credential?
No. The person answers for their own workspaces and personal items; company credentials and production destinations follow the organisation's rules.

More in Trustwelt AI Security

The other modules, and what works with this one.

Start with Trustwelt AI Security.

Agent gateway comes with Trustwelt AI Security, sold on its own per person per month. An organisation account is free.