Skip to main content

Any agent. Any model. Your rules.

The governed agent operating system.

Everything your AI agents need to do real work, in one place: identity, memory, tools and any model, under rules you set. These docs cover the four kinds of agent: Co-workers (one per employee), organisation agents (shared AI teammates with their own name in Slack), assistants (Claude, ChatGPT and Copilot connecting over MCP) and product agents (any framework, through the SDK, API or MCP). Every tool call they make through CortexHub follows one policy and lands on one audit trail.

Your AI app

Claude, Cursor, Windsurf, VS Code, or your Python agent runtime connects over MCP.

MCP + OAuth

CortexHub gateway

One endpoint. Search, policy checks, approvals, and audit on every call.

Governed invoke

External tools

Slack, Notion, GitHub, Jira, HubSpot, Stripe, and other integrations your workspace enabled.

Start with the job you need to do

Org admin

Create an Org Agent

Create a shared agent, give it one governed toolkit, and install it as a named Slack bot.

Admin quickstart →

End user

Connect an AI app

Claude, ChatGPT, Copilot, and any other MCP client such as Cursor or VS Code. OAuth sign-in, then connect toolkits.

Get started →

Developer

MCP SDK and frameworks

OpenAI Agents, LangChain, Vercel AI SDK, CrewAI, and more. Register the agent, bind an API key, open a governed session.

SDK quickstart →

Security

Governance and audit

Policies, approvals, guardrails, and activity logs for AI tool access.

Security guide →

Platform

Run agents for every team

Assign agent groups, publish tools, require approvals, and inspect the audit trail.

Agent guide →
Compare all roles →

Connect or build

End user

MCP clients

ClaudeInteractiveClaude DesktopInteractiveChatGPTInteractiveCursorInteractiveVS CodeInteractiveGitHub CopilotInteractiveGooseInteractivePostmanInteractiveMCPJamInteractiveArchestra.AIInteractiveWindsurfZedClineContinueJetBrainsReplitLibreChatGeminiMicrosoft Copilot
Any MCP client that supports remote OAuth and Streamable HTTP can use the same endpoint. Clients tagged Interactive render CortexHub’s in-chat Connect, Consent, and Operator Review prompts inline (per the official MCP Apps support matrix).

Add the MCP server, sign in with your CortexHub account, and connect toolkits in Cloud. No API keys or developer setup.

  1. MCP URL: https://mcp.cortexhub.ai/v1/mcp
  2. Sign in when your app prompts you
  3. Allow access, then connect toolkits in CortexHub Cloud
Get started guide →

Developer

Python MCP SDK

OpenAI Agents SDKLangChainLangGraphAnthropic APIVercel AI SDKCrewAIMastraPython SDK

Register the agent in CortexHub Cloud, bind an API key to it, then open a governed MCP session with session(agent=…, end_user_subject=…).

PyPI version 3.2.0
from cortexhub import Cortexhub

cx = Cortexhub(api_key="cxh_key_EXAMPLE")
session = cx.session(agent="helpdesk-bot", end_user_subject="user-123")
# session.mcp.url + session.mcp.headers → your MCP client
SDK quickstart →
MCP endpointhttps://mcp.cortexhub.ai/v1/mcp
Interactive appsBrowser OAuth (Bearer token)
Programmatic SDKAPI key (X-CortexHub-Key) + X-CortexHub-End-User-Subject

Learn the platform

Concepts and glossary

Organisations, the four kinds of agent, toolkits, groups, the Brain, and meta-tools: explained for admins and developers.

Learn →

Workspace admins

Enable toolkits, configure group access, approvals, guardrails, and integrations in CortexHub Cloud.

Admin guide →

Integration guardrails

Do

  • Use https://mcp.cortexhub.ai/v1/mcp as the MCP server URL
  • Route tool calls through CortexHub meta-tools (cortexhub_search_tools, cortexhub_get_tool_schemas, cortexhub_invoke_tools, cortexhub_get_task when parked, cortexhub_request_consent when needed)
  • Use Cortexhub().session(agent=…) for programmatic MCP

Don't

  • Put API keys or vendor credentials in the agent or chat
  • Use the bare host without /v1/mcp
  • Bypass CortexHub when your workspace expects governed tool access