PDQ Connect MCP

Early Access: The PDQ Connect MCP is in Early Access. The available tools and setup steps may change as we keep building it out, and you may run into rough edges.

The PDQ Connect MCP (Model Context Protocol) connects Claude.ai (and other AI chat assistants) directly to your data in PDQ Connect. MCP is an open standard that lets an AI assistant securely connect to a tool like Connect and act on your behalf. Once it’s set up, you can ask questions in plain English and the assistant looks things up and takes actions in Connect for you, without clicking through the console.

For example, you could ask “Which devices are missing the latest Chrome update?” or “Deploy the 7-Zip package to the IT test group.” The assistant only does what your PDQ account already permits, and every action is recorded in your Audit Logs. Setup takes a couple of minutes per assistant, and you only do it once.

Quick Setup

Open Settings → Connectors → Browse → Search for PDQ Connect (direct link)
Complete the OAuth flow to connect to your PDQ organization.

Create a .vscode/mcp.json file in your workspace:

{
  "servers": {
    "pdq-connect": {
      "type": "http",
      "url": "https://app.pdq.com/mcp"
    }
  }
}

Open the Command Palette (Cmd+Shift+P / Ctrl+Shift+P) and run MCP: List Servers.
Start the PDQ Connect server and complete the OAuth flow when prompted.

Run this command in your terminal:

claude mcp add --transport http pdq-connect https://app.pdq.com/mcp

Use the /mcp command to list and manage the MCP servers you have installed, and use the /context command to understand the context token usage of your current session, including the number of tokens used by each MCP server that’s enabled.

Go to chatgpt.com/#settings/Connectors (requires login).
Click Add Connector and enter the URL: https://app.pdq.com/mcp
Complete the OAuth flow to connect to your PDQ organization.

Open Settings → Plugins →MCPs.
Click Add server, select Streamable HTTP and enter the URL: https://app.pdq.com/mcp
Complete the OAuth flow to connect to your PDQ organization.

Example prompts

Not sure where to start? Once your assistant is connected, try one of these. Expand a prompt and copy it.

Find devices missing an app

Show me devices without Chrome installed in PDQ Connect in an interactive table

See what packages are available

What Firefox packages are in PDQ Connect?

Build a shareable dashboard

Create a dashboard of devices without Firefox or Chrome in PDQ Connect that I can share with my IT Director

Deploy to devices that need it

Could you add Chrome to any device that doesn't have it?

Build a “Patch Tuesday impact” report (advanced)

A longer, copy-and-paste prompt that has your assistant pull live fleet data from PDQ Connect, cross-reference the latest Patch Tuesday CVEs from the web, and produce a self-contained interactive HTML report.

You are helping me build an interactive "Patch Tuesday impact" report for my PDQ Connect
organization. Use my connected PDQ Connect tools for all device data and the web for CVE data.
Be pragmatic and accurate. Never invent numbers; if you can't measure something, say so.
STEP 1 - GET THE CVEs
- Web-search the most recent Microsoft Patch Tuesday and its notable CVEs (count, how many
  Critical, any zero-days or actively-exploited flaws, and the headline RCE/EoP CVEs).
- Patch Tuesday is the second Tuesday of the month. If the newest one has no published CVE
  data yet, use the most recent release that does and tell me which month you used.
STEP 2 - MEASURE MY FLEET (exact counts, not estimates)
Use search_devices. The base device row includes os, os_full_name, os_version (build),
require_reboot, and last_checkin_time. Tips that keep this accurate:
- search_devices caps at 250 rows. To get an exact COUNT for any segment, run a filter with
  limit:1 and read the "total_count" field. That's the reliable way to size each group.
- Get these totals:
  * Whole fleet: total devices.
  * By OS: macOS vs Windows clients (os_full_name contains "Windows 11"/"Windows 10")
    vs Windows Server (os_full_name contains "Server"), and each server version.
  * Pending reboot (exact, very actionable): filter devices[require_reboot]=true
  * Stale / unreachable: filter  devices[last_checkin_time] relative more than 7 days ago
  * Patch level by build: the current cumulative shows up as the highest build revision in
    the fleet for each Windows family (the 4th number, e.g. 26100.XXXX). Count how many are on
    the latest revision vs behind. Find the top revision present, then count devices below it.
STEP 3 - SCORE AND PRIORITIZE
- Map the headline CVEs to my fleet (e.g. a kernel/TCP-IP RCE touches every Windows device;
  HTTP.sys and Defender flaws are server-weighted; RDP-client RCEs hit admin workstations).
- Build a remediation order by blast radius. Usually: reboot-pending servers first (the fix is
  already staged, just needs a restart), then any actively-exploited flaw on isolated hosts,
  then rolling the latest cumulative to behind clients, then chasing stale devices back online.
STEP 4 - BE HONEST ABOUT LIMITS
State these plainly in a methodology footer:
- Patch level is INFERRED from OS build revision, not per-CVE/KB state.
- require_reboot and last_checkin are exact (straight from Connect).
- The MCP does not expose installed-software inventory, so app CVEs (SharePoint, SQL, Office)
  can't be confirmed per device here; note they'd need the Connect REST API or a software report.
- Name which Patch Tuesday month the CVE data is from.
STEP 5 - OUTPUT: a single self-contained interactive HTML file
Include: a KPI row (devices behind, pending reboot, on latest build, stale), a fleet posture
bar, filterable/expandable CVE cards each showing my exposure count, a prioritized remediation
queue, a server breakdown, and the methodology footer. Use vanilla JS only, no external libs
except Google Fonts. No localStorage.
LIGHTWEIGHT PDQ BRANDING (apply to the HTML)
- Dark theme. Page background #131515, panels/cards #1F2121, hairline borders ~#33393a.
- Body text white at ~85% opacity; never pure black or pure white. Muted text ~#8a9290.
- Headers in Space Grotesk, body in Inter, both from Google Fonts.
  Fallbacks: Lato for headers, Roboto for body.
- Accent color: teal #0A95E6 (or a brighter mint #5CDBBF) for highlights, links, key numbers.
- Status colors: success #007D54, warning #FF9933, error/critical #ff6b5e.
- Put PDQ's signature rainbow bar as a thin top stripe: five equal-width segments left to right,
  #FF9933, #FFFF33, #99CC33, #A2FFD4, #0099CC.
- Left-align text. Bold for emphasis only. Underline only for links. Don't rely on color alone
  to convey meaning.
- Don't recreate the PDQ logo; a plain "PDQ Connect" wordmark set in Space Grotesk is fine.

Audit logs

Every action that changes something through MCP is recorded in your Audit Logs, right alongside the actions taken in the console. This gives you a complete history of what was done, by whom, and from where.

Actions that are logged include:

Deploying packages to devices or groups (create_deployment)
Creating a device group (create_group)
Creating or editing an automation (create_automation)

Read-only actions such as searching devices, groups, or packages are not written to the Audit Logs, since they don’t change anything in your organization.

To help you tell MCP activity apart from everything else, the Audit Logs now include Source and Client columns:

Source shows where the action came from - MCP, API, Console, or Connected app. Actions taken through an AI assistant show a Source of MCP.
Client shows which app made the request, for example Claude Desktop, Claude Code, or ChatGPT.

Connecting a new assistant is logged too, so you’ll see a “Connected MCP client” entry the first time each app authenticates to your organization.

App connections

The App connections settings page gives you one place to review and manage every AI assistant and app that’s connected to your organization through MCP. Whenever someone completes the OAuth flow to connect an assistant like Claude or ChatGPT, that connection shows up here.

From this page you can:

See which apps and assistants are currently connected to your organization.
Review who authorized each connection and when.
Revoke a connection you no longer want. Revoking immediately cuts off that app’s access - any further MCP requests from it are denied until it’s connected again.

Revoking a connection only removes that app’s access. It does not change the account’s own permissions or roles in Connect.

Details

Available MCP tools

We’re continuing to build out functionality and add more tools over time.

Tool name	Action
`search_devices`	Search and query devices
`search_groups`	Find device groups
`create_group`	Create a device group
`search_packages`	Find packages to deploy
`search_deployments`	Search and query deployments
`create_deployment`	Deploy packages to devices
`search_automations`	List and filter automations
`create_automation`	Create or update an automation
`search_vulnerabilities`	Find and rank CVEs affecting your org
`search_device_vulnerabilities`	Find CVEs detected on specific devices
`search_top_vulnerable_devices`	Rank devices by number of open vulnerabilities
`get_vulnerability_details`	Look up a CVE’s details and your org’s exposure
`deploy_vulnerability_remediation`	Deploy the recommended package to fix CVEs
`ignore_org_vulnerabilities`	Ignore or un-ignore CVEs across your whole org
`ignore_device_vulnerabilities`	Ignore or un-ignore CVEs on specific devices

Interested in a specific tool and/or workflow? Let us know.

Authentication and permissions

The PDQ Connect MCP server enforces the exact same permissions as the Connect console. When you authenticate via OAuth, the MCP server acts as you and inherits your role and permissions. It cannot grant any additional access.

The rule is simple: if you can’t do it in the console, you can’t do it via MCP.

This means:

Every action taken through MCP is checked against your account’s permissions, just as it would be in the console.
You can only see and act on the devices, packages, groups, and other resources your account already has access to.
If you lack permission to perform an action in the console (for example, deploying a package), that same action will be denied via MCP.
MCP is not a way to bypass or escalate permissions. It is simply another interface to the access you already have.

Troubleshooting

Authentication issues

Make sure you complete the OAuth flow when prompted.
Try disconnecting and reconnecting: look for a “Clear authentication” or “Disconnect” option in your tool’s MCP settings.

Mark Rood

Updated August 03, 2026 16:31

Quick Setup

Example prompts

Find devices missing an app

See what packages are available

Build a shareable dashboard

Deploy to devices that need it

Build a “Patch Tuesday impact” report (advanced)

Audit logs

App connections

Details

Available MCP tools

Authentication and permissions

Troubleshooting

Authentication issues

Related articles