Claude · MCP · Ai Agents

Should your own business app open inside Claude? A decision guide

Anthropic's 10 November update lets apps open inside Claude. This guide is for SME and mid-market teams that run their own systems. It covers how to choose between a directory connector, a tools-only MCP connector and an MCP App, and which guardrails to build in first.

Softpiper Solutions9 min read
On this page

On 10 November, Anthropic announced (opens in a new tab) that a connector can now open a full tool beside the Claude conversation, or at full screen. Claude can then make changes inside that tool, and the user approves its actions. Most of the coverage has been about the launch names: Asana, Canva, Dropbox, monday.com, Vercel and others. For most SMEs and mid-market teams, that is not the question that matters.

The question that matters is about your own systems. That means the ops tool your team works in all day, the customer portal you built a few years ago, or the reporting app finance checks every morning. Should those open inside Claude? If they should, how much should Claude be allowed to do there?

This guide covers three realistic options, a four-question test for choosing between them, and the guardrails we would insist on before any write action reaches production.

What changed on 10 November

According to Claude's announcement (opens in a new tab), users can now open a tool beside the conversation or at full screen and ask Claude to make changes directly, as long as the connector supports it. Users approve Claude's actions the same way they already approve a connector's actions. For tools like Confluence or Notion, users get the full editor, signed in to their own account, rather than a limited preview.

The feature is rolling out on the web and in the Claude desktop app on all plans. It is part of a new Claude experience that brings chat and Cowork together. The first tools to support it are Asana, Canva, Carta, Dropbox, monday.com, Smartsheet, Vercel and Webflow. Figma, Harvey and Shopify are also building integrations.

Several other features shipped the same week. The same 10 November update (opens in a new tab) added:

  • Claude Sheets and Claude Whiteboards, two new artifact types.
  • Claude Data (formerly Dashboards), which connects to tools like Snowflake and Salesforce.
  • Claude auto-connect, in beta on eligible Team and Enterprise plans.

Claude Meetings was announced the same day (opens in a new tab). It began rolling out in beta (opens in a new tab) on Team and Enterprise plans on 12 November.

How it works: MCP connectors vs MCP Apps

Connectors are built on the Model Context Protocol (MCP) (opens in a new tab). Claude Marketplace lists more than 9,000 connectors and plugins, including community-built ones. A standard connector exposes tools, meaning actions and data that Claude can call. The results come back into the conversation.

Interactive connectors add another layer. They run on MCP Apps (opens in a new tab), an open MCP extension that lets an MCP server bring its own interface into AI applications. Two details matter if you are planning a build:

  • An MCP App is not limited to Claude. MCP Apps is an open extension, so the interface you build is not tied to Claude.
  • Developer support is in public preview. App tools and embedded sites are available to developers now, but you should expect them to change.

In short, a tools-only connector lets Claude act on your system through the conversation. An MCP App brings your system's own screens into Claude, and Claude works in them while the user watches.

Three options for your own systems

Option 1: Use a directory connector

If the system you care about is a mainstream SaaS product, check the directory first. If a maintained connector already exists, your job is setup and policy, not engineering. This is the cheapest option and should be your default.

The limit is that directory connectors cover other vendors' products. They will not cover an internal tool your team built.

Option 2: Build a tools-only MCP connector

For your own system, the next step is an MCP server that exposes a small set of well-defined tools, such as looking up an order, fetching a report or drafting a record. Claude calls these tools, and the results come back into the conversation as text, tables or artifacts.

This works well when the conversation itself can be the interface. It is quicker to build, easier to secure and easier to audit, because every action is a separate tool call. This is where we would start most integrations and APIs for the systems you already run.

Option 3: Build an MCP App

An MCP App brings your system's own interface into the Claude window. This matters when people need to see a record, queue or layout while Claude changes it, and check those changes on screen before approving them.

It is more work, because you are building and maintaining an interface on a developer feature that is still in public preview. The benefit is that users stay in one place. And because MCP Apps is an open extension, your investment is not tied to one AI vendor.

A decision test: four questions

Ask these four questions about each system you are considering.

  1. Visibility: do users need to see the thing change? If the output is an answer or a short summary, tools-only is enough. If the output is a changed schedule, board, form or layout, an MCP App is worth the extra cost.
  2. Approval frequency: how often will a write need a human decision? Occasional writes can be approved inside the conversation. If someone will approve dozens of changes in one session, it is easier and safer for them to see the record next to the request.
  3. Interface needs: does your current interface show things that text would lose? Colour-coded statuses, validation messages and dependent fields are often where your business rules live. If plain text would lose that context, embed the interface instead of rebuilding it in chat.
  4. Permissions: who owns access control today? If your app already has clear roles and each user signs in individually, both options can respect that. If permissions are thin or people often share accounts, fix that first. A connector should never be the first serious access control your system gets.

As a rule of thumb:

  • Mostly reads, rare writes and nothing visual to show: build a tools-only connector.
  • Frequent writes that someone reviews on screen: consider an MCP App.
  • A mainstream SaaS tool that already has a maintained connector: use the directory connector.

Security and admin controls you get out of the box

The platform gives you a reasonable baseline. According to Anthropic's update (opens in a new tab):

  • Claude works within existing permissions, and admins can limit what each connector does. For example, a connector can be allowed to read Slack but not send messages.
  • On Enterprise plans, admins decide who can use which products and connectors, and can set granular spend limits.
  • Auto-connect makes connecting apps like Notion, Atlassian or Dropbox a one-time setup. It is in beta on eligible Team and Enterprise plans. It is convenient, but it means you need your connector policy settled before users get access.
  • Enterprise data is encrypted in transit and at rest. Audit logs and a Compliance API work with more than 100 security and compliance tools, including Splunk and CrowdStrike.

For interactive connectors, admins can restrict a connector's tools (opens in a new tab) the same way they do today.

These controls cover Claude's side. The MCP server is your side, and for a custom system that is where most of the risk is.

Guardrails we'd build in regardless

When we scope AI agents and automation on Claude, we treat these guardrails as non-negotiable for any connector that touches a business system:

  • Read-only by default. The first version only offers lookups and reports. Write tools are added one at a time, and each one has a named owner.
  • Approve every write. Approval in Claude is the minimum. If an action moves money, changes stock or contacts a customer, also require confirmation inside your own system. That way a misread prompt cannot complete the action on its own.
  • Narrow tool scopes. A tool that only marks an invoice as approved is better than one that can update any field on any record. Small tools are easier to restrict, test and explain to an auditor.
  • Per-user identity. The connector acts as the signed-in user. It should never use a shared service account with broad access.
  • Logging on your side. Log every tool call in your own system, including the user, inputs, outputs and approval. Keep this alongside whatever passes through Claude's audit logs and Compliance API.
  • A kill switch. Have one setting that turns off all write tools without needing a new release.

Where this fits: ops desks, data views and back-office workflows

The best early candidates have something in common: frequent, structured work where a person already reviews each step.

  • Ops desks. Order, ticket or dispatch queues where staff check status and make small updates. Start with read tools, then add tightly scoped status changes.
  • Data views. A question like "which accounts slipped this month, and why?" suits a talk-to-your-data assistant, built as read-only tools over your reporting layer.
  • Invoices. In an invoice-to-Tally workflow, Claude extracts the data and drafts the entries, and a person approves each posting. This is a strong case for an MCP App, because the reviewer wants to see the entry before it is posted.
  • RFQs and purchase orders. RFQ and PO automation follows the same pattern: Claude drafts and compares, and a person sends.

If your system is a custom web app or platform, the connector can often reuse its existing API and roles. A modern codebase like Vyakarni, a React web app we built, is the kind of system where an MCP server can usually be added as a separate service without restructuring the app.

A 30-day path from pilot to rollout

Week 1: Choose and scope. Pick one system and one team, and run the four-question test. List five to ten read tools and no more than two write tools. Confirm that each user signs in individually.

Week 2: Build read-only. Ship a tools-only MCP connector with read tools, logging on your side and a kill switch. Use admin controls to limit it to the pilot group.

Week 3: Add writes behind approval. Turn on one write tool, with approval in Claude and confirmation in your system. At the end of the week, review every logged write with the person who owns the system.

Week 4: Decide. If users keep switching tabs to check what changed, scope an MCP App. If the conversation is enough, stay with tools-only and widen access.

When to wait

Hold off in any of these cases:

  • Your permissions are weak.
  • Nobody will own the connector after launch.
  • The use case depends on MCP App features, which are still in public preview for developers.
  • A maintained connector for the tool is likely to appear soon. Shopify, Figma and Harvey are already building integrations (opens in a new tab).

Next step

Softpiper is a Claude Partner Network · Select Services Partner, and we build on the systems our clients already run. If you have an internal tool, portal or reporting app and want to find out whether it belongs inside Claude, talk to Softpiper about a pilot connector. We'll start read-only.

Sources

Sources

  1. Your apps can now open in Claude, and Claude can work in them (Claude, Nov 10 2026) (opens in a new tab)
  2. More ways to work with Claude: Sheets, Whiteboards, and Interactive Connectors (Claude, Nov 10 2026) (opens in a new tab)
  3. Claude release notes (Claude Help Center) (opens in a new tab)
  4. Anthropic launches Claude Meetings to turn workplace discussions into plans and tasks (TechCrunch, Nov 10 2026) (opens in a new tab)

Researched and drafted with Pip, Softpiper's AI, and published by Softpiper Solutions.

Softpiper Solutions

AI software & automation studio. About us

Pip · Softpiper's AI

Ask Pip about this article

Turn the article into a plan for your business

Pip is Softpiper's AI. Ask how the ideas in this article apply to your stack, what we would build first, or for a tailored price range. A person from the team is one message away.