Who can use this feature?
The Chili Piper MCP server also works inside Microsoft Copilot Studio, using the same endpoint and the same API-key authentication described in How do I connect Chili Piper via MCP?. This article covers the Copilot Studio–specific setup steps, a gotcha in the connection wizard that isn't obvious the first time, and the limitations you should plan around.
Read the limitations before you build on this. Copilot Studio caps the tool list at 70 (the org MCP endpoint exposes more), and the connection periodically asks you to re-verify credentials mid-session. Neither is a misconfiguration on your end — see Limitations below before you commit a production workflow to it.
Table of Contents
Setup
The server URL is the same one used for every MCP client:
https://fire.chilipiper.com/api/fire-edge/v1/org/mcp- In Copilot Studio, open your agent → Tools → Add a tool → New tool → Model Context Protocol.
- Server name — anything you like.
-
Server URL —
https://fire.chilipiper.com/api/fire-edge/v1/org/mcp - Authentication — API key.
- Type — Header.
-
Header name — enter exactly
Authorization. Just that word — nothing else goes in this field. - Click Create, then Add tool → Create a new connection.
- The connection dialog is where the actual credential goes. Enter:
Bearer YOUR_API_KEY— capital B, a single space, then your key.
Generate the API key from Command Center → Integrations → Credentials → API Access Tokens (see Get your API key for the full steps).
Gotcha: the wizard's Header name field wants the word
Authorizationonly — theBearer <key>value goes in the connection dialog that appears one step later, not the wizard. If you pasteBearer <key>into Header name by mistake, Copilot Studio shows the error "Enter the complete server path to continue" — pointing at the Server URL field. The URL isn't the problem; the header name is. This is the single most common way to get stuck here.
Copilot Studio's MCP support uses streamable HTTP, which is what this endpoint serves — no extra configuration needed on that front.
Limitations
70-tool cap
Copilot Studio truncates the tool list at 70. The org MCP endpoint exposes more than that, so some tools are silently dropped and which ones is not something you control. In practice this means the agent may tell a user it can't do something it actually supports, or reach for a worse-fitting tool. We don't currently know whether the truncation is deterministic (e.g. first-70-in-server-order) or something else.
Recommendation: don't build one general-purpose agent against this connector. Build narrow, purpose-built agents so each stays under the cap — for example:
- Reporting agent — meeting volume, breakdowns, exports
- Troubleshooting agent — single meeting lookup, CRM sync checks, lookup by guest email
- Routing agent — distributions, rules, routers (admin-only; this is where you'd put write access if you want it)
Reconnect prompts
The connection will periodically interrupt a request with a prompt to open the connection manager and re-verify credentials, with a Retry button. This isn't a misconfiguration and recurs even after you've verified once. If you're piloting this with a customer or on a live call, mention it up front so it doesn't read as a broken integration.
Authentication notes
Use an API key, not OAuth, for this connector:
- Third-party MCP servers have a documented pattern of failures with Copilot Studio's OAuth dynamic client registration flow in other integrations. We have not tested whether the Chili Piper MCP hits this.
- Even where OAuth connects, Copilot Studio has been reported to not silently refresh tokens for MCP connections.
- Chili Piper's OAuth path is Admin-only on paid accounts; API keys can be scoped per user via the Permissions dialog, which is a better fit for a Copilot Studio deployment anyone on the team might use.
Environment and governance: MCP servers in Copilot Studio run on Power Platform connectors, so any Data Loss Prevention (DLP) policy governing Power Platform connectors also governs this connection — and DLP enforcement for Copilot Studio has been on by default tenant-wide since early 2025. Microsoft's own guidance is to avoid the Default environment, since tenant-level DLP policies commonly restrict custom connectors there, and to use a Developer or Sandbox environment instead. This has been confirmed working in a Dev environment; whether it works in a Managed or production environment depends on that tenant's Power Platform admin policies and hasn't been verified by us. Treat a production rollout as a separate conversation with the customer's Power Platform admin.
Running MCP tools also requires Generative Orchestration (not Classic Orchestration), which carries its own licensing and Copilot Credit implications — a trial can build and test an agent but can't publish it to production.
Where this works
"Copilot" spans several distinct Microsoft products, and only one of them can reach a custom MCP server like this one:
| Surface | Works with this MCP server? |
|---|---|
| Copilot Studio (full) | ✅ Yes — this is the supported path |
| M365 Copilot chat | ❌ No — can't reach arbitrary third-party MCP servers |
| Copilot Studio for Teams / lite agent builder | ❌ No — limited to Classic Orchestration; MCP requires Generative Orchestration |
| GitHub Copilot in VS Code | ✅ Yes, full MCP support — but this is a developer tool, not a business-user agent builder |
Skills in Copilot Studio
Our GitHub-hosted Chili Piper Skills install directly in Claude and ChatGPT, but the install mechanism — the plugin marketplace — is specific to those clients and doesn't exist in Copilot Studio.
What does carry over is the content: each Skill is a documented, step-by-step query recipe, and an agent's Instructions field in Copilot Studio is designed to hold exactly that kind of procedural guidance. Porting a Skill's steps into an agent's Instructions text is a natural next thing to try.
This is on our radar as a follow-up exploration, not something we've validated yet — we don't currently know whether Copilot Studio's orchestrator follows a multi-step recipe as reliably as Claude does. If you try it before we publish guidance, we'd like to hear how it goes.
Comments
0 comments
Article is closed for comments.