The governed agent operating system

Let AI agents act in your systems. Keep control of every action.

CortexHub checks every tool call before it runs. Routine work goes straight through, risky actions wait for the people your policy names, and anything your policy forbids is blocked. Every decision is signed into evidence your auditor can verify.

  • Claude, ChatGPT, your own agents and ours, through one gateway
  • One sentence per policy, checked by code, not an AI model
  • Signed evidence mapped to 7 compliance frameworks

Any agent. Any model. Your rules.

Policy decisionsAcme Inc, this morning
Live policy
  • ClaudeClaudeAllowedHubSpotHubSpot: Read the Q3 pipelineDecided by: Toolkit decision: HubSpot read tools run
  • Ava, a Co-workerNeeds approvalStripeStripe: Refund $1,200Decided by: Org floor: refunds over 500, Finance approves
  • ChatGPTChatGPTNeeds approvalGmailGmail: Email 240 customersDecided by: Group floor: sends externally to over 100 recipients, the member approves
  • GitHub CopilotGitHub CopilotBlockedGitHubGitHub: Delete a protected branchDecided by: Tool override: delete_branch on protected branches is blocked
  • Support agent (SDK)Support agent (SDK)BlockedZendeskZendesk: Export every customer recordDecided by: Org floor: exports data that touches PII, blocked
  • CursorCursorAllowedJiraJira: Open a ticketDecided by: Toolkit decision: Jira write tools run

A toolkit the agent's group was not given? The call is refused.

The problem

Agents now move money, email customers and change records. Who approved that?

Claude, ChatGPT and the agents your teams build connect straight to company systems. Each one arrives with its own keys, its own settings and its own logs, so there is no single place to decide what they may do and no single record of what they did. When an agent gets it wrong, the liability is yours.

No single control

Every assistant and agent is set up on its own. A rule you set in one does not reach the next.

No record you can stand behind

When a customer or an auditor asks what an agent read, changed or sent, the answer is spread across vendor logs and screenshots.

Security has to say no

Without control and proof, agents stay in pilots, or people use them anyway, outside any policy.

$670K

added to the average breach cost at organisations with high levels of shadow AI.

IBM, Cost of a Data Breach Report 2025
40%

of agentic AI projects, at least, will be cancelled by the end of 2027, due to rising costs, unclear value or inadequate risk controls.

Gartner, 25 June 2025
One gateway

Every agent goes through one gateway. Your policies sit outside all of them.

Claude, ChatGPT, GitHub Copilot and other assistants connect over MCP, the open standard AI assistants use to reach tools. Agents your developers build connect through the CortexHub SDK. CortexHub's own Co-workers (an AI assistant for each employee) and organisation agents (shared AI teammates for a team) use the same path. Before any call runs, the gateway checks who is acting, for whom, and what the call would do.

A registry or GRC tool
The CortexHub gateway
Where it works
Beside your AI: an inventory and a risk score
In the path of every tool call, before it runs
What it can stop
Nothing at runtime. It reports after the fact
Allows, holds for approval or blocks each call
Which agents
The ones someone remembered to register
Every agent connected through the gateway, Claude and ChatGPT included
Evidence
Questionnaires, attestations and screenshots
A signed record of every decision, verifiable offline

Control sits outside every agent, ours included, so no agent can talk its way past it. Credentials stay in a vault and never reach the model. See what happens when something goes wrong

Policies

Build each policy as one sentence. Code decides every call, never an AI model.

Set an organisation floor every group inherits, a floor for each group, and one decision for the read, write and destructive tools of each toolkit. Tighten a single tool with conditions on its own fields. What the page shows is what runs.

  • One sentence per policyPick what the tool does, the decision, who it is accountable to and when it applies: "When a tool that moves money as a refund then require approval, accountable to Finance, only when money over 500."
  • Deterministic, checked by codeEvery floor item, toolkit decision and tool override is checked by code, the same way on every call. No AI model decides a live call.
  • New tools are covered automaticallyCortexHub classifies every tool by what it does (moves money as a payment or a refund, sends externally, deletes) and recommends floors for the tools you connect. Floors match what a tool does, not its name, so a toolkit you connect next month is covered the day you connect it. A call no decision covers is blocked.
  • Tighten one toolAn override on one tool asks first or blocks it, only when its own fields match, such as an amount over 500. An override only tightens its toolkit decision, never loosens it.
  • Try a call before it happensEnter sample arguments for a tool and see what your policies decide for that group, who approves and what decided it. Nothing runs.
PoliciesAcme Inc
Org floor

When a tool that moves money as a refund then require approval, accountable to Finance, only when money over 500.

Moves money as a refund (money over 500)Needs approval
Applies to
Every group
Governs now
StripeShopifyand any refund tool you connect later
Approves
Finance

Amounts are read in real units: a refund a tool sends in cents is compared with 500 pounds or dollars.

Dry run, Support group
Stripe refund 1,200.00
Needs approvalFinance approves. Decided by: Org floor, moves money as a refund (money over 500).
Approvals

Risky actions wait for the right person.

When a policy says a person must decide, the call pauses before it runs, and the people that policy is accountable to decide: the member the agent works for, your back office, or a group or people you name, such as Finance. An org admin can step in, recorded as an override. They decide in Slack, in the Approvals inbox or inside their AI assistant. Scheduled work nobody is watching waits too.

Follow one request from start to finish. Pick a kind of agent to replay it.

Akhil asks a Co-worker in Slack to update a deal in HubSpot.

SlackAsked in Slack1 / 12

Akhil messages the Co-worker in Slack, where the team already works.

Evidence

Every action, signed into evidence your auditor can check.

Each decision is recorded with the agent, the person it acted for, the policy that decided and who approved. Records are signed and hash-chained, so any change breaks the chain and names the record. The evidence maps to SOC 2, ISO/IEC 42001, the EU AI Act, NIST AI RMF, NIST AI 600-1 (the GenAI profile), ISO/IEC 23894 and the OWASP Top 10 for LLM applications.

  • SOC 2
  • ISO/IEC 42001
  • EU AI Act
  • NIST AI RMF
  • NIST AI 600-1 (GenAI profile)
  • ISO/IEC 23894
  • OWASP Top 10 for LLM apps

Who, what and why

Every record names the agent, the person or end user it acted for, the policy that decided and the approver.

Verifiable offline

Export the evidence for any period and verify it without trusting CortexHub.

Gaps point to their fix

Readiness is measured from what your agents actually did, and every gap names the setting that closes it.

This is readiness evidence for your own audits. CortexHub is CASA validated, and SOC 2 is in progress. See the framework mapping

Proof

Control you can check for yourself.

Every call is decided by code, the evidence verifies offline, and an agent reaches only what its group was given.

7
compliance frameworks your evidence maps to
No AI judge
every live call is checked by code, never by a model judging it
Least privilege
an agent reaches only the toolkits its group was given, and anything else is refused

Every governed action is signed into a tamper-evident, hash-chained record.

FAQ

Common questions

CortexHub is the governed agent operating system: one gateway that every AI agent's tool calls pass through, with your policies applied before anything runs. Routine work goes straight through, risky actions wait for the people your policy names, and anything your policy forbids is blocked. Every decision is signed into tamper-evident evidence mapped to 7 compliance frameworks. It covers Claude, ChatGPT, GitHub Copilot and other MCP assistants, agents built on any framework through the CortexHub SDK, and CortexHub's own Co-workers and organisation agents.
No. Teams keep the assistants they already use and connect them to the CortexHub gateway over MCP. CortexHub governs the tool calls those assistants make through the gateway and records them on the same audit trail, so the same policies apply whether a call comes from Claude or from a CortexHub Co-worker.
Registry and GRC tools govern from above: they inventory and score your AI, but they do not sit in the path of a tool call, so they cannot stop one. CortexHub governs at the point of action: every call is allowed, held for approval or blocked before it runs, and every decision is signed into evidence. It also gives you governed Co-workers to deploy under the same policies.
On Policies you build each floor item as one sentence: what the tool does, the decision, who it is accountable to and when it applies. Each group adds its own floor and one decision for the read, write and destructive tools of each toolkit, and you can tighten a single tool with conditions on its own fields. CortexHub classifies every tool by what it does and recommends floors for the tools you connect; you decide which to use. Everything is checked by code, the same way on every call, never by an AI model. Policies match what a tool does, not its name, so new toolkits are covered automatically, and a call no decision covers is blocked. A dry run on any tool shows what would happen and why, before a real call.
The agent receives a clear response explaining the action was blocked by policy. Nothing runs. The attempt is recorded in the audit trail, so your security team can see exactly what was tried, by whom, and when. An agent reaches only the toolkits its group was given, and a call to anything else is refused.
Actions your policy marks as needing approval are paused before they run, and the people that policy is accountable to decide; an org admin can step in, recorded as an override. When it is the member the agent works for, they decide in the chat, in the approval card their AI assistant shows where it supports one, or in Slack, with the Approvals inbox as the backstop. When it is your back office, or a group or people you name such as Finance, the approval goes to the Approvals inbox, a Slack channel or a signed webhook to your existing tools.
CortexHub maps its readiness evidence to SOC 2, ISO/IEC 42001, the EU AI Act, NIST AI RMF, NIST AI 600-1 (the GenAI profile), ISO/IEC 23894 and the OWASP Top 10 for LLM applications. Records are signed and hash-chained, so an auditor can verify offline, from an export alone, that nothing was changed after the fact. This is readiness evidence for your own audits. CortexHub itself is CASA validated, and SOC 2 is on our roadmap and in progress.

Start with one workflow

Your first governed workflow, live in minutes.

Connect an assistant, connect a system, set a policy and run it. Tell us what you want your agents to do and we will help you choose where to start.

For the enterprise

See it govern your agents.

Book a demo to see policies, approvals and the audit trail on your own tools and security requirements.