Skip to content
KiweelyKiweelyKiweely

Connecting Claude or Copilot to Workday via MCP: Seven Security Questions to Answer First

Connecting Claude or Microsoft Copilot to Workday’s MCP server routes every tool call through Agent Gateway and Agent System of Record security. Before going live, verify Delegate vs. Ambient ASU modes, Available To interaction policies, token lifetimes, and audit logs.

  • By Kiweely
  • 7 min read
On this page

Connecting Anthropic Claude, Microsoft Copilot Studio, Cursor, or Google Agentspace to Workday’s Model Context Protocol (MCP) server does not bypass Workday security, because every JSON-RPC call must pass through Workday Agent Gateway and an Agent System User (ASU) registered in the Agent System of Record (ASOR). However, misconfiguring execution modes, skill interaction policies, or OAuth token handling can easily grant unintended background access or break production sessions every 24 hours.

Before your security administrator signs off on wiring an external assistant to a tenant, walk through these seven verification questions in Workday.

1. Does Your Tenant Meet the Licensing and Regional Gateway Prerequisites?

Workday’s native MCP server (https://{region}.agent.workday.com/mcp) is not an open endpoint on every Core HCM tenant. Per Workday’s external agent documentation, connecting a registered external agent via Workday MCP requires Workday Extend Professional along with Agent System of Record enabled in the tenant. Your contract must also be on the Universal Main Subscription Agreement (UMSA), or—if your organization remains on a legacy Main Service Agreement (MSA)—have Innovation Services and AI data contributions explicitly enabled.

You also need to pin the exact regional Agent Gateway host matching the data center that hosts your tenant (us.agent.workday.com, eu.agent.workday.com, uk.agent.workday.com, ca.agent.workday.com, au.agent.workday.com, sg.agent.workday.com, in.agent.workday.com, or jp.agent.workday.com) and pass the unique wd-agent-tenant-alias header on gateway calls. Pick the right regional endpoint from day one: changing regions later resets your historical agent analytics in Agent Management Hub.

2. Are Skills Registered in Delegate Mode or Ambient Mode?

This is the most consequential security decision in ASOR. When an external agent is registered via POST /asor/v1/agentDefinition (requiring Setup: Agents in the Agent System of Record functional area) and configured in Agent Management Hub, each skill runs in one of two execution modes—and Workday can create up to two separate ASUs per agent to isolate them:

Security Control Delegate Mode (Interactive User) Ambient Mode (Autonomous Background)
Security ControlRuntime Identity Delegate Mode (Interactive User)Intersection of signed-in Workday user and Delegate ASU Ambient Mode (Autonomous Background)Ambient ASU only (no human user in the loop)
Security ControlOAuth Mechanism Delegate Mode (Interactive User)OAuth 2.0 Authorization Code (On-Behalf-Of user login and Allow/Deny consent screen) Ambient Mode (Autonomous Background)Unique X.509 public/private keypair with JWT Bearer assertion (iss = Client ID, sub = ASU username)
Security ControlPermission Boundary Delegate Mode (Interactive User)User must pass the skill’s Available To policy and hold domain/BP permissions for the tool’s APIs Ambient Mode (Autonomous Background)Determined solely by domain/BP policies granted to the generated Ambient Agent Security Group
Security ControlAudit Trail (View AI Agent User Audit Trail) Delegate Mode (Interactive User)Logs the ASU as By User and the employee as On Behalf Of User Ambient Mode (Autonomous Background)Logs only the Ambient ASU as By User

For interactive assistants like Claude Desktop or Microsoft Copilot, always use Delegate mode. If a developer registers an interactive MCP integration using an Ambient skill to avoid building an OAuth login redirect, every employee chatting with that bot shares the Ambient ASU’s unconstrained permissions—erasing row-level supervisory boundaries and hiding which human prompted the action in View AI Agent User Audit Trail. Note also that ASUs never support Workday proxy accounts or worker delegation.

3. Do Both the Interaction Policy and Tool Domain Policies Align?

In Delegate mode, Workday enforces a strict two-gate check on every MCP tool invocation:

  1. Agent Interaction Policy (Available To): In Agent Management Hub > Configure Agent, the user must belong to at least one security group listed on the Available To prompt for that enabled skill. If they are missing from Available To, Workday blocks the interaction immediately—even if the user holds full domain access to run the underlying task in the Workday UI.
  2. Underlying API & Field Security: Once the user clears Available To, Workday evaluates their standard domain and business process security policies (down to field-level access) against the Workday IDs (WIDs) of the SOAP, REST, Graph, WQL, or RaaS operations bound to that skill.

Never guess whether those two layers line up. Open the View Security for Agent Skill report (secured to Manage: Agents) for your registered agent and inspect the Functional Status column for every tool:

  • Full Match: The security group in Available To holds all required domain/BP permissions.
  • Review Required: The group holds partial permissions; tool calls may fail mid-conversation or omit fields.
  • Config Required: One or more required security policies have zero security groups assigned.
  • Not Evaluated: Available To is blank, meaning nobody can invoke the skill.

Use the Test My Setup button right inside that report (requires Security Administration in the System functional area) to trace exact policy pathways for a test user before testing in Claude or Copilot.

4. How Does Your Client Protect Secrets and Handle 24-Hour Refresh Expiry?

When you complete Configure Agent for an external agent the first time, Workday displays the OAuth Client ID and (for Delegate skills) the OAuth Client Secret once. That screen never displays again, and the credentials point to the ASU record visible on the All AI Agent Accounts report (Workday Accounts domain).

Watch out for a common developer shortcut shown in local testing tutorials: pasting a raw Bearer token into a static local config file.

json
{
  "mcpServers": {
    "Workday MCP": {
      "url": "https://us.agent.workday.com/mcp",
      "headers": {
        "Authorization": "Bearer <hardcoded-asu-token>"
      }
    }
  }
}

That snippet is fine for a 10-minute smoke test with {"jsonrpc": "2.0", "id": 1, "method": "tools/list"}, and dangerous everywhere else. Plaintext config files on developer laptops leak into dotfiles and backup drives, and Workday’s Delegate ASU access tokens expire in at most 4 hours (14,400 seconds) (60 minutes in some flows), while refresh tokens expire after 24 hours. Production MCP clients must implement full OAuth 2.0 Authorization Code flow with secure OS keychain or vault storage, re-prompting user sign-in cleanly when the 24-hour refresh window closes. Also ensure your token parser accepts both string and integer types for expires_in, as Workday’s test and production endpoints can return either type.

If you are synchronizing agents from Microsoft Copilot Studio or Azure AI Foundry into ASOR via the Create Microsoft Agent Sync Configuration task, the trust boundary extends into Microsoft Entra and Power Platform. Microsoft Agent Sync requires application permissions with Microsoft Entra Global Administrator consent, a dedicated Service Principal mapped to a Power Platform Application User, and explicit PowerShell connector role assignments (Set-AdminPowerAppConnectorRoleAssignment) on Dataverse custom connectors. Scope that Service Principal solely to the specific Copilot Studio environment IDs approved for Workday—never your entire tenant-wide sandbox sprawl.

Inside Workday, review the Passport tab on each external agent’s profile in Agent Management Hub (View Agent Passport Security for Agents). Passport stamps record third-party security assessments, assessor identities, and grant dates so governance teams have an auditable record inside the tenant.

6. What Stops a Runaway Agent Loop from Hitting MCP Throttles or Burning Flex Credits?

Autonomous reasoning models retry failed tool calls and chain pagination queries aggressively. Workday enforces explicit platform guardrails on Workday Agent-Ready MCP Tool Calls:

  • Concurrent Request Limit: 2,000 active requests
  • Burst Limit: 500 requests
  • Refresh Rate: 500 requests

Beyond HTTP 429 Too Many Requests throttling, external API and agent gateway traffic on UMSA contracts with signed platform policies feeds into your Platform Consumption Console meters. And remember one operational detail during incident response: if an external agent misbehaves and you click Deactivate in Agent Management Hub, status propagation to Agent Gateway and MCP Gateway takes up to 1 minute—and Workday never auto-deactivates agents when your Flex Credit balance hits 100%.

Keep Narrow Skill Definitions

Do not bundle 40 read and write REST/SOAP WIDs into a single catch-all MCP agent. Split read-only lookup skills from mutating business process skills so you can restrict write skills to narrow Available To security groups and deactivate a noisy skill without taking down employee lookups.

7. Are You Solving Employee Self-Service or Consultant Configuration Work?

Before investing weeks into Extend Professional licensing, ASOR agentDefinition payloads, and OAuth token brokers, confirm you are solving the right problem. Workday’s MCP server architecture is built to expose curated runtime business tools—like checking PTO balances, submitting expenses, or looking up org charts—to employees inside enterprise chat clients. It does not expose Workday configuration tasks without public APIs, build calculated fields, edit Workday Studio .lar assemblies, or validate Workday Extend app bundles.

If your real goal is giving Workday consultants, HRIS admins, and Extend developers an AI agent that inspects tenants, configures business processes, builds reports and EIBs, develops Extend and Studio projects, and runs browser tests—without publishing OAuth brokers or mixing client credentials—compare how Kiweely Vine works. Kiweely runs natively on macOS and Windows with strict per-Client credential isolation and explicit consent on every change; review our security architecture and MCP tool testing lessons before deciding which path fits your team.

Frequently asked questions

What license is required to connect an external agent to Workday's MCP server?

Workday's MCP server connection for registered external agents requires Workday Extend Professional, Agent System of Record (ASOR) setup, and either a Universal Main Subscription Agreement (UMSA) or Innovation Services enabled under a Main Service Agreement (MSA).

How does Workday enforce user permissions when Claude or Copilot calls an MCP tool?

In Delegate execution mode, Workday checks two layers at runtime: first, whether the signed-in user belongs to a security group listed in the skill's Available To interaction policy, and second, whether that user's own domain and business process security policies permit the underlying API and field access.

How long do Workday Agent System User OAuth tokens last for external agents?

For third-party external agents using the OAuth 2.0 Authorization Code flow in Delegate mode, ASU access tokens expire in up to 4 hours (14,400 seconds) and refresh tokens expire after 24 hours. Workday does not currently allow customizing access token expiry for third-party agents.

Can you test external MCP agents in Workday using Start Proxy?

No. Workday's agent security framework explicitly blocks both proxy access and delegation for Agent System Users (ASUs). You must test Delegate-mode agent interactions using real non-Production test accounts assigned to the target security groups.

Your Workday work, done by AI.

Kiweely Vine is the AI agent that builds, deploys and tests Extend apps as real people, makes changes in any Workday task screen, answers from your real tenant and builds reports.

Get the next one in your inbox.

One email a month for Workday teams: what’s changing in Workday and what to retest, one tip tried in a real tenant, and what’s new at Kiweely.

We’ll email you once to confirm. Unsubscribe in one click from any issue.