Skip to main content
Quint records what AI agents do on your company’s machines and in your containers, and lets your security team set rules for them. For each agent session, Quint keeps two separate records: what the agent said it was about to do, and what the machine then saw it do. Where the two disagree, Quint raises a finding and names the risk. Quint has three parts. An endpoint runs on developer machines. A container integration covers agents that run in CI jobs, servers and sandboxes. The dashboard is where your security team investigates, writes rules and exports evidence. Developers keep working as before.

The problem

AI coding agents such as Claude Code, Cursor and Codex run with the same access as the developer who starts them: the repository, the shell, credentials on disk and the network. They read files, run commands and call tools on their own, many times a minute. The controls most companies already run each see part of this:
  • An AI gateway sees the traffic routed through it, and nothing that happens on the machine afterwards.
  • EDR sees operating-system activity but has no record of what the agent meant to do, so it cannot tell an expected action from an unexpected one.
  • The agent’s own logs report what the agent says it did. An agent misled by a prompt injection or a tampered tool can report something other than what ran.
None of them compares what the agent declared with what ran. Credential theft, data leaving the machine and changes nobody asked for show up in that gap.

How it works

Endpoints and containers send each agent’s declared actions and observed activity to Quint. Quint compares the two, checks each action against your rules, and sends signed rules back to the machines. Your team sees the results in the dashboard, in approval requests, in alert channels and in your SIEM.
  • Developer machines. Macs on Apple Silicon, Windows and Linux are supported. You roll the endpoint out through the MDM or configuration tool you already run: Microsoft Intune, Jamf, Mosyle, Iru (formerly Kandji), Ansible, Puppet or Chef. See Platform support.
  • Containers. Agents in CI jobs, servers and sandboxes run as Quint workloads in Docker or on Kubernetes.
  • Without the endpoint. Quint can read from endpoint security tools, SaaS agent platforms, enterprise AI assistants, cloud AI platforms and your identity provider. See Agentless coverage.

Main use cases

Key capabilities

  • Declared versus observed. Quint captures what the agent declares (through the agent’s hook, its model traffic and its telemetry) apart from what the machine records (process, file and network activity). Findings are named by risk in eight categories: credential access, exfiltration, persistence, execution, unexpected change, intent mismatch, approval bypass and tool supply chain. See Investigate a finding.
  • Agent coverage. Claude Code, Codex and Cursor are recorded in full, and an armed rule can stop them at the tool call. Gemini CLI is recorded in full. Ten more agents, including GitHub Copilot CLI, Qwen Code and OpenHands, are recorded at the tool-call level, and any other agent at the network level. See Supported agents.
  • Rules with shadow mode. Rules are versioned and signed, and a person reviews each set before it is published. Each rule records, asks or refuses. Every ruleset starts in shadow mode, where it records what it would have done against the last 14 days of real traffic. You arm enforcement one group of machines at a time. See How policies work.
  • Approvals. A rule set to ask holds the action for a person. Approvers see secrets masked and must reveal them before they can approve. When they deny, they can send the agent a sentence of guidance. See Decide an approval.
  • Sentinel. Your team asks questions about the fleet in plain English. Sentinel answers from your workspace’s own data, shows every query behind each figure, and drafts rules for a person to review and publish. See Ask Sentinel.
  • Investigation and incidents. Each finding opens onto the session timeline, the process lineage, child sessions and the health of the machine’s capture. Findings take verdicts, and incidents move from Open to Closed.
  • Audit trail. The audit trail covers agent activity and every action a person takes in the console. You can check its chain on request and export it. See Audit Trail.
  • Integrations. Alerts and approval requests go to Slack, Microsoft Teams, PagerDuty or a signed webhook. Events and the audit log go to Splunk, Datadog, Elastic or OpenSearch, Microsoft Sentinel, Sumo Logic, Amazon S3 or a generic webhook. See Integrations.
  • Privacy controls. Quint masks the credentials it recognises before upload. Connections from programs that are not agents are discarded on the machine, and the contents of connections are never read. A policy setting turns content upload off. Each workspace’s data stays in one data region and is kept apart from every other workspace’s. See What Quint records.

Business value

Next

Last modified on November 13, 2026