What’s new with Microsoft 365 Copilot declarative agents
Microsoft 365 Copilot declarative agents just got three significant upgrades, and they landed close enough together that I’m rebuilding parts of the upcoming Voitanos workshops around them. If you’ve been building declarative agents, you’ve hit the same walls I have: text-only interactions, knowledge sources limited to existing M365 content, and a dev loop that meant bouncing across portals, JSON files, and editors to wire anything together. MCP Apps, embedded knowledge, and the new Microsoft 365 Agents Toolkit plugin just knocked all three down. Here’s what changed and why it matters.
Until very recently, declarative agents were limited to natural language interactions, a narrow set of knowledge sources (SharePoint, OneDrive, a handful of URLs, Teams chats, and Copilot connectors), and a dev loop that required portal-hopping to wire anything together. If you need a primer on how Copilot processes those interactions end-to-end, start with How Microsoft 365 Copilot Works. All three of those constraints just fell away. Let’s walk through them.
How to add interactive UI to declarative agents with MCP Apps
MCP Apps let Microsoft 365 Copilot declarative agents render interactive UI — expense approval forms, project dashboards, document comparisons — directly inside the Copilot chat experience. Microsoft quietly shipped this capability a couple of months ago but didn’t announce it until April 7, 2026:
Learn how MCP Apps transform agent capabilities in Microsoft 365 Copilot, enabling seamless app interactions.
https://devblogs.microsoft.com/microsoft365dev/mcp-apps-now-available-in-copilot-chat
Here’s what’s going on. The Model Context Protocol (MCP) recently added a way to return not just data from an MCP server to an MCP client, but a full user experience alongside that data. There are two flavors, and both work with Copilot:
- MCP Apps: An extension to the existing MCP spec from Anthropic, who originally released the Model Context Protocol a couple of years ago.
- OpenAI Apps SDK: OpenAI’s own take on the same idea, released around the same time.
What “interactive UI” actually means in practice is that your agent can drop a real interface into the chat, like an expense approval form, a project status dashboard, or a side-by-side document comparison. It’s not a card or a link. It’s a live surface the user can interact with.
There are two surface modes you need to know about:
- Inline mode: A lightweight widget that shows before the model’s response. This one is required if you go this route. It’s great for previews, confirmations, and quick decisions, like approving a pull request or any other approval flow where you want a bit more context than a chat message and a button.
- Side-by-side mode: A fully immersive workspace in the main area of the Copilot experience, with the chat pushed to the margin. Your user can still talk to the agent through natural language, but they also get a real UI to work in for multi-step editing, complex layouts, or comparisons.
For declarative agent developers, this means you’re no longer stuck with text. The agent can hand the user a real UI to act in and then carry the action forward from there.
Here’s what I really appreciate about how Microsoft did this. The underlying MCP server model means your existing functionality, authentication, and integrations don’t have to change. The UI components layer on through the meta property, which is additive and backward compatible. That’s a good design call from Microsoft for once, and it means you can retrofit existing MCP-backed agents without rewriting everything.
Launch partners already shipping these include Adobe Express, Coursera, Figma, monday.com, and the Outlook compose and scheduling experience. By mid-April, these should start showing up in the Microsoft 365 Agent Store.
Learn more about MCP apps in Microsoft 365 Copilot from the following resources:
Learn how MCP Apps transform agent capabilities in Microsoft 365 Copilot, enabling seamless app interactions.
https://devblogs.microsoft.com/microsoft365dev/mcp-apps-now-available-in-copilot-chat
Learn how to add interactive UI widgets to MCP server-based declarative agents.
https://learn.microsoft.com/microsoft-365/copilot/extensibility/declarative-agent-ui-widgets
Interactive UI applications that render inside MCP hosts like Claude Desktop.
https://modelcontextprotocol.io/extensions/apps/overview
Learn how to use Apps SDK by OpenAI. Our framework to build apps for ChatGPT.
https://developers.openai.com/apps-sdk
Embedded knowledge for self-contained declarative agents
Declarative agents now support embedded knowledge — up to 20 files uploaded directly into the agent as knowledge sources, no SharePoint or OneDrive required. Document files (.doc, .docx, .pdf, .ppt, .pptx, .txt) support up to 512 MB each; spreadsheets (.xls, .xlsx) are capped at 30 MB.
| File types | Max per agent | Max size per file |
|---|---|---|
| .doc, .docx, .pdf, .ppt, .pptx, .txt | 20 total | 512 MB |
| .xls, .xlsx | 20 total | 30 MB |
Before this landed, if you wanted to use knowledge in your agent (and you almost certainly did), you were working with SharePoint and OneDrive files, public URLs with tight limits, Teams chats, Copilot connectors, and a handful of other M365 sources. For real-time data, you’d reach for REST APIs through actions or MCP servers.
Now you can ship a fully self-contained agent with its reference content baked in. No SharePoint provisioning gymnastics. No waiting on IT to set up a site. No separate deployment of content to your customers that they might accidentally modify. Everything the agent needs to answer questions lives inside the agent itself.
To prioritize embedded content over Copilot’s foundational knowledge, enable “Only use specified sources” in the Configure tab of your agent manifest. If you want the agent to lean on the content you shipped with it first, that’s the switch to flip.
A few things to know going in. Anyone with access to the agent has access to the embedded content, so there’s no Information Barriers support here. Sensitivity labels propagate, meaning the highest-priority label on any embedded file becomes the agent’s label. This isn’t supported in GCC Moderate. And don’t try to embed DKE files, files with user-defined permissions, password-protected files, or cross-tenant encrypted files, because they’ll fail silently or break agent creation entirely.
For the full list of supported knowledge sources, see the official docs:
Learn about the different types of knowledge sources that you can add to your declarative agent when you build your agent with the Agent Builder feature in Microsoft 365 Copilot.
https://learn.microsoft.com/microsoft-365/copilot/extensibility/agent-builder-add-knowledge#embedded-file-content
Microsoft 365 Agents Toolkit plugin for Work IQ
The Microsoft 365 Agents Toolkit plugin lets you scaffold, configure, and deploy a full declarative agent — including an MCP server, auth setup, widget rendering, and deployment — through natural language prompts in Claude Code or GitHub Copilot CLI.
This is the change I’m personally most excited about, because it fixes the developer experience problem I’ve been complaining about for a long time. Microsoft shipped this as part of the microsoft/work-iq repository called microsoft-365-agents-toolkit@work-iq. If you’ve worked with Claude Code or GitHub Copilot CLI, you already know the plugin format. It’s a common way to package a collection of skills, agents, MCP servers, and (in the case of Claude Code) slash commands into your project.
To install the plugin, run these three commands:
# register the marketplace (if you haven't already)
/plugin marketplace add microsoft/work-iq
# install the plugin
/plugin install microsoft-365-agents-toolkit@work-iq
# restart the coding assistant...
# or reload the plugins if using Claude Code
/reload-plugins
Once it’s installed, you can literally say things like “Create a declarative agent that uses an MCP server with a dashboard widget for my insurance claims” and it scaffolds the entire thing for you.
The plugin includes two skills:
| Skill | What it handles |
|---|---|
| Declarative Agent Developer | Scaffolding, manifest authoring, API plugin wiring, MCP server wiring, capability configuration, deployment |
| UI Widget Developer | MCP servers with rich chat widgets using MCP Apps and OpenAI Apps SDK |
Why is this meaningful? This is the first time you can go from “I have an idea” to “deployed agent with a working UI widget” without context-switching across portals. That’s the dev loop I’ve been begging for, and it’s the kind of experience I already love in other agentic development environments.
One pragmatic note. This plugin is still in preview, things will change, and you need a tenant admin to consent to it. I wouldn’t bet a production deployment on it just yet, but it’s absolutely worth starting to prototype with.
MCP Server and CLI for accessing Work IQ. Contribute to microsoft/work-iq development by creating an account on GitHub.
https://github.com/microsoft/work-iq/tree/main/plugins/microsoft-365-agents-toolkit
What I’m pulling into the upcoming workshops
All three of these are going into the next round of Voitanos workshops with hands-on demos. But I’m also adding two things that aren’t new and that I keep teaching because they unlock so much value when you combine them with everything above.
The first is TypeSpec for declarative agent development. TypeSpec lets you define your agent manifest in a strongly-typed DSL and compile it to the JSON schema Copilot expects, eliminating an entire category of manifest authoring errors. It’s faster, less error-prone, and much easier to refactor than hand-editing JSON. I’ll show how it folds into the new toolkit plugin flow.
The second is Microsoft Graph as the swiss army knife. Here’s the trick most people miss: you can do a huge amount with your M365 data inside a declarative agent that DAs don’t natively support, by leaning on Microsoft Graph through an API plugin. Need to pull calendar events, list recent emails, or query a user’s manager chain? All of that works through an API plugin backed by Microsoft Graph. And here’s the kicker. You don’t have to write any code to do it. I’ll demo it end-to-end, including the full SSO setup.
The combination of these five capabilities is, in my opinion, the new floor for what a serious declarative agent developer should know:
- MCP Apps — interactive UI inside Copilot chat
- Embedded knowledge — self-contained agents with baked-in reference content
- Microsoft 365 Agents Toolkit plugin — agentic dev loop for scaffolding and deployment
- TypeSpec — typed DSL for agent and plugin manifests
- Microsoft Graph via API plugins — no-code access to M365 data with SSO
Where to go next
Want to learn more? Join me! I’ve got a few places where I’m walking through all of this live over the next few weeks:
Live webinar - A Developer’s Guide to Declarative Agents
Wednesday, April 15, 2026, 12 - 1:30p ET
I’m going to walk through everything in this article and demo the highlights. If you stick around to the end, I’ll share a special discount code for my upcoming virtual workshop, good for a limited time. You have to be registered to get it.
Discover the pro-developer path to extending Microsoft 365 Copilot with declarative agents — built in VS Code with the Microsoft 365 Agents Toolkit.
https://www.voitanos.io/webinars/microsoft-365-copilot-developers-guide-declarative-agents-20260415/
Last week’s webinar - Evaluate Your Microsoft 365 Copilot Agent Options
If you’re not sure which agent creation approach is right for your scenario (Copilot Studio, declarative agents, agent builder, SharePoint agents, or Microsoft Foundry agents), start here:
Read the webinar recap:
Microsoft 365 Copilot supports four agent creation options: SharePoint agents, Agent Builder, Copilot Studio, and declarative agents built with the Microsoft 365 Agents Toolkit (ATK). Each sits on a spectrum from simple to powerful, and the tool you choose locks you into a specific orchestrator that directly affects query results and document retrieval.
https://www.voitanos.io/blog/microsoft-365-copilot-evaluate-your-agent-options-webinar-recap-20260408/
… or watch the webinar recording:
Learn to extend Microsoft 365 Copilot with custom agents. Compare SharePoint Agents, Copilot Studio, and Declarative Agents to connect external data sources.
https://www.voitanos.io/webinars/microsoft-365-copilot-evaluate-your-agent-options-20260408/
3-day virtual workshop - Building Declarative Agents for Microsoft 365 Copilot
April 28-30, 2026
This is where everything in this article gets taught hands-on. I’ve updated the curriculum to include MCP Apps, embedded knowledge, and the new toolkit plugin. Early bird seats are already gone, but group pricing is available if you have five or more developers on your team ( contact us for details).
Get the most from your Microsoft 365 Copilot investment with declarative agents, bringing data from LOB systems, and apply instructions for unique scenarios.
https://www.voitanos.io/workshop-microsoft-365-copilot-build-declarative-agents/
Full-day workshop at the Microsoft 365 Community Conference (April 19-24, 2026)
April 20, 2026 · Orlando, FL
I’m running a full-day workshop on Monday of the conference. Learn how to create Declarative Agents with me! Attendees also get a deep discount to my virtual workshop and access to the recordings.
Join me in Orlando, Florida in April 2026 for the Microsoft 365 Community Conference to get the most value from your Microsoft 365 Copilot investment workshop!
https://www.voitanos.io/blog/joinme-m365conf-orlando-2026/
European Collaboration Summit (May 5-7, 2026)
May 5, 2026 · Cologne, Germany
In early May, I’m co-delivering a full-day workshop with my good friend Mark Rackley comparing Copilot Studio agents with declarative agents. Attendees also get a deep discount to my virtual workshop and access to the recordings.
Join me in May in Cologne for ECS 2026! Learn how to get the most value from M365 Copilot when when to use Copilot Studio or using Visual Studio Code!
https://www.voitanos.io/blog/joinme-ecs2206-cologne-2026/
Wrapping up
The declarative agent story has genuinely changed in the last few weeks. MCP Apps give you real UI inside chat. Embedded knowledge lets you ship self-contained agents with their reference content baked in. And the new M365 Agents Toolkit plugin for Work IQ finally gives us the agentic dev loop we’ve been waiting for. Pair those three with TypeSpec and Microsoft Graph, and you’ve got a toolkit that lets you build things that would’ve been painful or impossible a couple of months ago.
Pick one of the three new capabilities and try it this week. Which one feels most relevant to what you’re building right now? Drop a comment below, leave a reaction, or ask a question. I’d love to hear what you’re planning to build first.
Microsoft MVP, Full-Stack Developer & Chief Course Artisan - Voitanos LLC.
Andrew Connell is a full stack developer who focuses on Microsoft Azure & Microsoft 365. He’s a 22-year recipient of Microsoft’s MVP award and has helped thousands of developers through the various courses he’s authored & taught. Whether it’s an introduction to the entire ecosystem, or a deep dive into a specific software, his resources, tools, and support help web developers become experts in the Microsoft 365 ecosystem, so they can become irreplaceable in their organization.
