Summary: On January 26, 2026, Anthropic launched interactive connectors and MCP Apps so tools such as Slack, Canva, and Figma can render live interfaces inside Claude. MCP Apps also became the first official MCP extension, allowing an MCP server to provide rich UI to multiple supporting clients. The design uses sandboxed iframes, auditable JSON-RPC messages, and consent controls for UI-initiated tool calls.
January 26, 2026 is the day tools like Slack, Canva, Figma, Box, and Clay stopped behaving inside Claude like APIs that hand back text. With interactive connectors and MCP Apps, Anthropic turned them into interfaces that open up within the conversation — visible, operable, and live. Have Claude draft a Slack message and you can look it over before it posts; hand it a rough paragraph and a FigJam flowchart comes back — none of it requires leaving the chat window.
What gives the launch its actual heft is the companion post published the same day on the Model Context Protocol blog: MCP Apps has become the first official MCP extension, so any MCP server can now ship connectors carrying a rich user interface — and the destination is no longer Claude alone. The chat window is turning from a question-answer surface into an application runtime, and for the first time a cross-vendor open standard sits underneath that idea.
Interactive Connectors and MCP Apps, Explained
The old connector model was a round trip with no face on it: Claude read from and wrote to external services over APIs, and whatever came back landed in the conversation as text or files — no interface to look at, nothing you could reach in and adjust. Interactive connectors flip the arrangement. External tools appear inside the conversation as interactive UI components, and every step the model takes sits in plain view, editable as you go. Anthropic’s stated rationale is blunt — “analyzing data, designing content, and managing projects all work better with a dedicated visual interface.”
Each app hands Claude a logged-in instance of the service behind it. In practice that means Claude can search Slack conversations and shape a message, rework a deck’s branding in Canva while you watch, or page through document previews in Box. TechCrunch reports the feature reaches paid plans first; to switch an app on, look for the “interactive” label in the featured section of claude.ai/directory.
Nine Apps at Launch, and What Each One Does
The opening lineup runs nine apps deep and is almost entirely workplace territory:
- Amplitude: put together analytics charts, dig into trends, tune parameters on the fly
- Asana: convert a conversation into projects, tasks, and timelines
- Box: find files, preview documents, lift out the pieces that matter
- Canva: draft presentation outlines and apply branding as it happens
- Clay: dig into companies and contacts, fetch funding data, write outreach drafts
- Figma: turn text and images into flowcharts and Gantt charts inside FigJam
- Hex: ask questions of your data and get interactive charts, tables, and citations back
- monday.com: refresh boards, hand out tasks, chart progress
- Slack: look through conversations, draft and format messages, give them a once-over before posting
Salesforce has already said it will come in through Agentforce 360, carrying enterprise context with it. The thread worth pulling is Cowork, the general-purpose agent tool built on Claude Code that shipped January 12: when these apps were announced, Cowork could not touch them, and Anthropic marks the integration “coming soon.” An agent that goes and updates a Figma marketing graphic or reads from the company’s Box account belongs to the next chapter of this playbook, not the current one.
Under the Hood: MCP’s First Official Extension
The technical proposal landed in November 2025, resting on the open-source MCP-UI work and the OpenAI Apps SDK, with MCP maintainers, OpenAI, and the MCP-UI community shaping it together. The machinery, reduced to its moving parts:
- A tool announces its UI metadata through a _meta.ui.resourceUri field
- The server serves the UI resource under the ui:// scheme, bundled HTML and JavaScript inside
- The host picks it up, renders it in a sandboxed iframe, and exchanges JSON-RPC with the server via postMessage
On the code side, the official SDK ships as the npm package @modelcontextprotocol/ext-apps, organized around an App class whose methods include app.connect(), app.callServerTool(), and app.updateModelContext(). An app built on it can log events, open links in the user’s browser, send follow-up messages, and even push updates back into the model’s context. The example servers in the repo — threejs, map, pdf, system-monitor, sheet-music — remain the quickest on-ramp for anyone building their own.
Not Just Claude: One Build, Every Client
If one part of the MCP Apps announcement decides how this plays out, it is the client list. Claude (web and desktop), Goose, and VS Code Insiders support it today; ChatGPT began rolling it out that same week. JetBrains, AWS (Kiro), and Google DeepMind (Antigravity) have each said they will follow. OpenAI shipped its own Apps system back in October 2025, but that one lives inside a walled garden; the MCP take is an open standard across clients — write the app once and, in principle, every supporting host can render it. The endorsements came from both camps: David Soria Parra, MCP co-creator at Anthropic, said he is excited about what MCP Apps opens up, while OpenAI’s Nick Cooper said “we’re proud to support this new open standard.” Once Microsoft, JetBrains, AWS, and Google have all put their names on the same UI protocol, the interface-layer standards fight in the agent ecosystem is effectively settled.
How the Security Model Works, and the Lesson for Product Teams
Third-party UI living inside a conversation raises the stakes, which makes the security design load-bearing. MCP Apps confines all UI to sandboxed iframes, insists on pre-declared templates that hosts can review, keeps every JSON-RPC message auditable, and can hold a UI-initiated tool call until explicit user consent arrives. The guidance Anthropic wrote into its own Cowork docs deserves to be copied straight into your playbook: don’t grant permissions you don’t need, be cautious with sensitive material like financial documents, credentials, or personal records, and consider a dedicated working folder for the agent instead of broad filesystem access.
For product teams the read is direct: where AI products compete is moving from “model answers” to “app execution.” If your product already runs an MCP server, the next question is which operations have earned an upgrade to interactive UI — the interface itself is now part of the protocol.