Home / Agent Hub
Cimento Agent Hub
Keep AI agents inside policy.
Agent Hub checks every shell command and MCP call from Cursor, Claude Code, and the Codex CLI against your policy, inside the agent, before it runs. Each finding comes with the evidence behind it.
The Problem
A document can’t stop a tool call.
01
Agents act unobserved
They read files, run shell commands, call tools, and connect to MCP servers: thousands of actions a day, most unseen by the developer who started the session.
02
Text steers them
A paragraph of untrusted input from a cloned README or a fetched page can redirect an agent holding a developer’s credentials, source tree, and shell.
03
Metadata isn’t enough
Your MDM can list which agents are installed. But a shell command looks routine until you see the key pasted into it.
How It Works
Enforcement inside the agent.
Checks before it runs
The Cimento guard runs in each agent’s own hook and checks shell commands, MCP calls, and other supported tool calls against a cached policy, on the device, in milliseconds.
Tells the agent why
A block carries a reason the agent reads, so it can find another way to finish the task. The person gets a nudge in Slack explaining the rule.
Explains every finding
Each finding names the person, agent, and device, links to the session steps that show what happened, and suggests a next step when it needs your call.
The Baseline
Starts soft. Escalates on repeat.
Cimento’s baseline watches from day one, and you choose when it enforces. Turn enforcement on and a rule set to block nudges for its first seven days. After that, within any 30 days, a person’s first match nudges, the second blocks, and the third blocks and alerts your team.
Default · Block
Unapproved MCP servers
An agent calls an MCP server that isn’t approved for the person’s team. Approve a server once per team, tied to its command or URL where Cimento can resolve it, not just its name.
Default · Block
Secrets and credentials
A known key format appears in a shell command or a tool’s arguments. The key is masked in Cimento’s record, and once the soft start ends, later matches are stopped before they run.
Default · Fix
Cimento guard installed
The guard is missing, removed, or overridden. Agent Hub re-adds a missing user-level guard in Claude Code and Cursor; otherwise the exposure stays open until your MDM fixes the setup.
Privacy by Design
Evidence, not an archive.
Full content is what lets you tell a routine command from one with a pasted key. It’s also a liability, so Agent Hub keeps as little of it, for as short a time, as the job allows.
Opt-in: content capture stays off until your company turns it on
Masked on the device: known secret formats are replaced before anything is buffered
Behind a permission: only admins you authorize can read session content
Short retention: content from a session without a finding expires after a week
No outside models: explanations come from Amazon Bedrock in Cimento’s AWS account, never OpenAI or Google
Ships through your MDM: signed packages, no kernel extension, outbound HTTPS only
FAQ
Which AI coding agents does Agent Hub cover?
Will it slow developers down?
What happens when we turn enforcement on?
Does Agent Hub replace our EDR or DLP?
How do we deploy it?
Measure first. Then enforce.
Agent Hub watches from day one and enforces when you’re ready, across Cursor, Claude Code, and Codex.
