VCF Agentic AI Hub 0.2.0 is a virtual appliance that lets you ask questions about a VMware Cloud Foundation lab in plain English, and draft PowerCLI and Python scripts. Broadcom ships it as a Fling. The release notes, dated 6 November 2026, describe it as an experimental tool for lab and non-production use only, with no SLA or warranty.
This release adds live NSX, vSAN and flow agents alongside the Script Assistant. I haven't installed 0.2.0 yet, so this post is based on the release notes and the user manual.
What VCF Agentic AI Hub is
MCP is the Model Context Protocol. It's how the hub's agents, and outside tools such as Claude Code, talk to the appliances you connect. The hub itself is a single VM on VCF Management Services, Broadcom's Kubernetes platform, with Keycloak handling sign-in.
There are two roles. vcf-administrator has full control of the hub: users, endpoints, the language model and diagnostics. vcf-viewer is read-only on the hub, and the manual recommends it for anyone who uses the agents but doesn't need to configure the platform. That doesn't mean everything the hub does on NSX is read-only, though. I've given that its own section further down.
You work in the built-in chat, and the hub picks an agent based on your question, or you can choose one yourself. Claude Code and Codex can also connect to the hub's MCP endpoint and sign in with a hub account. The release notes say the vSAN server isn't available to those outside clients.
What 0.2.0 can do
NSX queries
The NSX Query Agent is read-only. You ask in normal language about inventory, how objects are connected, whether a policy object has realised, and statistics such as DHCP pool usage. It doesn't cover NSX Federation, runtime state such as gateway high availability or group members, or questions about what isn't connected. On NSX 4.2.x, it only searches policy-mode resource types.
Datapath, BGP, capacity and upgrades
The VCF Networking Assistant is the troubleshooting side. It checks whether two VMs or IPs can reach each other, finds where and why traffic is dropped, looks at BGP neighbours, and checks Edge and Virtual Network Appliance capacity, including the last 24 hours of alarms. Datapath Troubleshooting can give you a diagram or a full report.
Four upgrade skills cover a go or no-go check, the version path, the upgrade order, and where to download the bundle. They're for planning only. They won't start an upgrade, troubleshoot one that has already failed, or download the bundle for you.
The assistant doesn't change firewall rules, routing or other network configuration. It does run Traceflow, Livetrace and the NSX upgrade precheck, though, and it installs diagnostic runbooks on NSX Manager as it goes. The section on what it will and won't change explains why that matters.
vSAN
The vSAN Agent checks cluster health, disks, objects, storage policies, capacity and performance on both ESA and OSA, and explains what it finds with knowledge base references. The manual describes this as inspection and diagnosis. It doesn't change the cluster, and if it gives you guidance for a procedure, that doesn't mean it can run the procedure for you.
It needs vSAN 9.0 or later, and vCenter and ESXi 9.0.0 or later. The release notes say the vSAN MCP server is only for this built-in agent. You can add Log Insight if you want it to read more logs. When your question is about vSAN, select the vSAN agent in the chat.
Flows and the Script Assistant
The VCF Operations for Networks assistant answers questions about flows, such as counts, top talkers, drops and anomalies. It's read-only, and it doesn't change VCF Operations for Networks, NSX or vCenter. It only works with flow data, up to 30 days of it, grouped by two properties at most, and you can't export the flows from the chat. Heavier per-flow work runs in the Python Sandbox, which is on by default, and the raw flows aren't sent to the model.
The Script Assistant is still there. It searches the official PowerCLI and VCF Python SDK documentation and returns the syntax and examples, so Claude Code, Cursor, Visual Studio Code or Codex can draft a script. It doesn't run the script. The manual's example is still a PowerCLI script that powers off a VM by name. You get a script to review, and the hub doesn't power the VM off.
What it will and won't change
On the hub itself, vcf-viewer is read-only. A viewer can't configure MCP servers, endpoints or the language model. An administrator can.
On NSX, the hub doesn't change firewall, routing or other network configuration. The exception is the diagnostic runbooks and the upgrade precheck. The hub doesn't count those as configuration changes, so:
- a viewer can start them
- the runbooks install on NSX Manager without a prompt
- they run as the NSX account you added when you onboarded NSX, not as the person who asked
The manual says that account should be Enterprise Admin. A lesser NSX role still answers inventory and health questions, but some skills come back partial or fail. Either way, the NSX audit logs show the onboarded account, not the hub user.
The vSAN checks and the flow assistant don't change their products, and the Script Assistant only reads documentation.
What you need
The installer is a Go binary for Linux, macOS and Windows, on both amd64 and arm64. The manual lists macOS Tahoe/Golden Gate and Windows 10 Enterprise, but no Linux version. The release notes put the package at about 17GB.
For the appliance, you need:
- vCenter 8.0 or later.
- Room for one VM with 20 vCPUs, 40GiB of RAM and 426GB of disk.
- Four free static IPv4 addresses: one for the bootstrap VM, the platform VIP, the node IP, and one external address for the hub. The bootstrap VM is created for the install and removed if it succeeds.
- Forward and reverse DNS records for the hub's FQDN, on the external address.
- Optionally, a signed certificate. If you skip it, you get a self-signed certificate to export at the end.
IPv6 isn't supported, and neither are DHCP or .local names.
The agents also need a language model. The manual supports Gemini, through Google AI Studio or Gemini Enterprise Agent Platform, or an OpenAI-compatible endpoint. The release notes say the built-in agents use gemini-3.8-flash. If you only want to host the MCP servers and don't need the chat agents, you can leave the model unset.
For NSX, the manual says the query agent and the networking assistant need 4.2.4 or later, or 9.1.1 or later. The release notes list a wider set of versions for the NSX MCP server itself: 4.2.2, 4.2.3, 4.2.4, 4.2.5, 9.0.0, 9.0.1, 9.0.2, 9.1.0 and 9.1.1. VCF Operations for Networks needs 6.14.x, 9.0.x, 9.1.x or later, and it only works with a local account.
Run the installer
Unzip the download. On macOS, double-click the .pkg. When it says the installation was successful, that only means the installer app is on your Mac. A terminal and a browser tab then open so you can carry on. On Windows, double-click the .exe, and on Linux, run the binary from a terminal. The manual's command for the web wizard is:
./vcf-ai-installer wizard --mode=webThe first screen is a checklist covering vCenter reachability, the four static addresses, DNS, the VM size and the optional certificate. You can also import a configuration file from an earlier run. When you're ready, select I have what I need - Continue, then work through these screens:
- vCenter credentials. Enter the server, username and password, then connect. The screen says the password is never saved to the browser.
- vCenter resources. Pick the datacenter, cluster, datastore, network, resource pool and VM folder.
- Network. Enter a pool of at least four static IPv4 addresses, plus the gateway, prefix, DNS, a primary VIP, the internal cluster CIDR, NTP and an SSH password hash.
- Ingress. Enter the platform FQDN. The platform VIP is worked out from that name, but you do enter an instance VIP. Add your own certificate, key and CA, or leave them empty.
- Bootstrap password. Set a password of at least 15 characters, with a lowercase letter, a digit and a special character. There's a switch for SSH password login and a timeout that defaults to 30 minutes. You can leave the appliance root password blank, and the screen says one is generated for you and shown once at the end.
- Review. If you want a file to reuse in case the install fails, select Generate Config Script. Then install.
The log stays on screen while it runs, and the manual warns it can take a while. If a run fails after it has created VMs, don't just start it again. Delete the bootstrap VM and the management node from the resource pool and folder the installer created first. The manual's example VM names look like "VCF Agentic AI Hub-0816".
Installing from the same subnet as the appliance is supported in this release.
Set up users and appliances
Add a viewer
Open https://<your-vcf-ai-fqdn> and sign in as admin with the bootstrap password. Go to Settings > User Management > Local Users, add a user and assign VCF Viewer. The manual's example user is jsmith.
Copy the temporary password before you close the dialog. It's only shown once, and the user has to change it the first time they sign in.
To change it, the viewer signs in, opens the user menu and selects Change Password. Viewers can see Agents and Authorizations under Settings, but they can't configure the hub.
Connect your appliances
Before the agents are any use, an administrator has to turn on the MCP servers that aren't already on, and add each appliance under Settings > AI Configurations with its address and an account.
Whatever that account can do is the limit for every hub user, viewer or administrator. For NSX, use Enterprise Admin if you want the precheck and the runbooks. For VCF Operations for Networks, a local account that can read the API is enough, on HTTPS port 443.
The release notes say the VCF Operations for Networks, NSX, vSAN, Script Assistant and Log Insight servers aren't enabled by default, but the Python Sandbox is. One part of the manual says the VCF Operations for Networks server is on by default, so check its toggle after you install.
Connect Claude Code and Codex
The MCP endpoint is https://<your-vcf-ai-fqdn>/mcp. Enable the servers you want to use before you connect.
Claude Code
For a lab with the self-signed certificate, the manual turns certificate checks off. That means you're not verifying who you're talking to, so don't do it against a certificate you don't trust.
export VCF_AI_URL="https://<your-vcf-ai-fqdn>"
export NODE_TLS_REJECT_UNAUTHORIZED=0
claude mcp add -t http vcf-ai "$VCF_AI_URL/mcp"
claude mcp login vcf-aiSign in as the hub user in the browser, and claude mcp list should show the server as Connected. Inside a session, /mcp or /mcp vcf-ai starts the same login.
On a machine without a browser, Claude Code v2.1.191 and later can log in headlessly:
claude mcp login vcf-ai --no-browserOpen the URL on another computer, sign in, and paste the whole callback address back into the prompt. The shell needs to be bash or zsh.
Codex
Codex needs CLI version 0.148.0 or later, or 0.156.0 or later for headless login. Sign-in has to use Dynamic Client Registration. Client ID Metadata Document sign-in doesn't work.
codex mcp add vcf-ai --url "$VCF_AI_URL/mcp" --oauth-client-registration dcr
codex mcp login vcf-ai --oauth-client-registration dcrFor headless Codex, add --no-browser to the login command. The callback you paste is hidden as you type. codex mcp list should then show the server as enabled.
Collect a support bundle
Sign in as admin, go to Settings > Support Bundle and select Generate. Retention is a whole number of hours from 1 to 8. When the status shows Successful, download the bundle from the actions menu. If it fails, the error is on the status badge.
The release notes also point to per-request traces, which the owner of a chat session can open or download. An administrator can't open someone else's traces. You can also rename, pin and delete sessions.
If you report a bug on the Broadcom Community Flings board, include the appliance version and either a support bundle or a session trace.
Upgrade to a new release
There's no in-place upgrade. You deploy a new appliance and import an encrypted backup, and the installer for the release you're moving to handles both the export and the import.
Limitations
It's a Fling, so keep it in the lab. Features can change before any generally available release, and there's no support contract behind it.
A few limits are easy to miss:
- Datapath Troubleshooting does a shallower job on VLAN-backed VPC segments, and VPC distributed transit gateways aren't fully supported.
- Edge and VNA capacity checks look at up to 10 nodes, and you only get a recommendation if there's been a datapath CPU or memory-pool alarm in the last 24 hours.
- With more than one NSX endpoint, some requests use the default one instead of the one you named.
- Flow analysis reads at most 20,000 flows for the detailed pass.
- Upgrade help is planning only.
The runbook behaviour is the one I'd brief a team on. A viewer can get a diagnostic runbook installed on NSX Manager, with no prompt, under the onboarded Enterprise Admin account. That's a change on NSX, even if the hub doesn't call it one, and even though firewall and routing rules aren't touched.
Where to download it
0.2.0 is on the Broadcom support portal, under Flings. The package is about 17GB.
VCF Agentic AI Hub 0.2.0 on the Broadcom support portal
Resources
- VCF Agentic AI Hub 0.2.0 download (Broadcom support portal, Flings)
- Broadcom Community, Flings, where the release notes ask you to report bugs and ideas
Feedback
If you found this guide useful, subscribe to the mailing list or follow me on X and YouTube. You can contact me through any of these with issues or suggestions for this guide.
Questions? Email me (contact@virtualg.uk) or DM me on X (@virtualg_uk).