Claude connectors extend what Claude can do by linking it to approved outside tools, data sources and services through the Model Context Protocol (MCP). Whether you are judging connectors for your organisation, building your own or setting rules for teams, this guide explains how they work, what they can and cannot do, and how to use them responsibly.
Availability note: Which connectors you can use, and what they can do, may depend on your Claude plan, the app you use, your region, admin settings, the outside app, your permissions and Anthropic’s current product setup. The features described here come from public information and may change.
Claude Enterprise Training → Claude Code Training What is MCP?Quick Answer
Claude connectors are connections built on the Model Context Protocol (MCP) that let Claude use approved outside tools, pull in context from connected services and, where allowed, take actions in those systems. Claude only gets the access each user already has in the connected service. Organisation admins decide which connectors are switched on and what they may do. Which connectors you can use depends on your plan, the app you use, your region and Anthropic’s current product setup.
What Are Claude Connectors?
By default, Claude works with what is in your current conversation and any files or documents you give it. Connectors add structured connections to approved outside systems that respect permissions, so Claude can pull in relevant context, refer to live or indexed data and, in some cases, take allowed actions in connected apps.
Connectors are built on Model Context Protocol (MCP), an open protocol that defines how AI systems communicate with external tools and data sources. Claude.com hosts a connector directory of pre-built integrations. Organizations can also build custom connectors by implementing remote MCP servers.
A key principle: connecting a service does not give Claude open access to that system. Claude gets only the access the person asking already has in the connected app. Organisation owners on Team and Enterprise plans can also limit what actions each connector may take.
Important: Which connectors you can use, and what they can do, depends on your Claude plan, the app you use, your region, admin settings, the outside app, your permissions and Anthropic’s current product setup. Not every connector is available on every plan or app.
Five Types of Claude Connectors
Anthropic lists several types of connector. Which ones you can use depends on your plan, the app you use and your admin settings.
ARemote / Web Connectors (Pre-built)
- • Cloud-hosted integrations available from the Claude connector directory.
- • Accessible across Claude's web interface, mobile apps, Claude.ai for Work and — where applicable — Claude Desktop and Claude Code.
- • Connect Claude to services such as project management tools, communication platforms, document systems and more.
- • Each connector has its own capability profile (read, write, interactive) — verify what each connector can actually do.
- • Administrator enables and controls which connectors are available to which users.
Desktop Extensions (Local)
- • Run locally on the user's machine — not in the cloud.
- • Available only in Claude Desktop and Claude Code. Not available through the web interface or mobile apps.
- • Provide access to local tools, files, development environments and machine-level integrations.
- • Suitable for developer workflows where local system access is needed.
- • Security model differs from remote connectors — local execution requires careful review of what the extension can access.
Custom Connectors via Remote MCP
- • Organizations or third parties build and host a remote MCP server that Claude connects to.
- • Enables integration with internal systems, proprietary data sources or specialized tools not in the pre-built directory.
- • Requires development resources, security review and ongoing maintenance.
- • Anthropic does not check the security of third-party MCP servers; organisations must check it themselves.
- • Available where remote MCP connectors are supported; plan and administrator configuration applies.
Claude Code MCP (Developer-Configured)
- • Developers configure MCP servers locally (per-project) or remotely within Claude Code.
- • Supports local MCP servers, project-scoped MCP config files and remote MCP server connections.
- • Primarily for developer workflows: connecting Claude Code to databases, APIs, version control, CI/CD or custom tools.
- • Configuration lives in project or user-level config files; scope and trust level are developer-managed.
- • See Claude Code documentation and training for detailed MCP configuration patterns.
API-Level MCP Connector (Application Integration)
- • Developers building Claude API applications connect their app to remote MCP servers.
- • Enables applications — not just end-user Claude interfaces — to use MCP-connected tools and data.
- • Useful for building products and internal tools where Claude serves as an AI layer over connected systems.
- • Requires Anthropic API access and development against the MCP specification.
- • The developer or organisation is responsible for setup and security in its own app.
Remote Connectors vs Desktop Extensions
Remote connectors and desktop extensions are built differently, are available in different places and carry different security risks. Check which type suits your use and platform.
| Aspect | Remote / Web Connectors | Desktop Extensions |
|---|---|---|
| Hosting | Cloud (Anthropic or third-party hosted) | Local machine |
| Interfaces | Web, mobile, Claude.ai for Work, Desktop, Code (where applicable) | Claude Desktop and Claude Code only |
| Web / Mobile availability | Yes | No |
| Use case | Cloud services, SaaS tools, organizational systems | Local files, development tools, machine-level access |
| Auth model | OAuth per-user, or enterprise-managed auth (where available) | Local config, developer-managed |
| Admin controls | Org owner can restrict or block actions | Developer-configured; org admin controls vary |
| Security review | Pre-built directory connectors reviewed by Anthropic; third-party MCP servers are not | Developer responsibility; local execution scope |
These design details come from Anthropic’s public product information and may change. Check what is currently available in your plan and setup.
The Connector Permission Model
Claude connectors operate within a layered permission model. An action is only permitted if it passes every layer. Understanding this model helps organizations design appropriate governance and helps users understand why certain requests may not be fulfilled.
Always Allow
The connector can take this action automatically when relevant, without per-request user confirmation.
Needs Approval
Claude presents the proposed action and the user must explicitly confirm before it is executed.
Blocked
The action is not permitted regardless of user request. Blocked actions cannot be executed through this connector.
Policy options for organisation owners on Team and Enterprise plans, depending on Anthropic’s current product setup. Check the current admin controls in your plan’s documentation.
Read Connectors vs Write Connectors
The Claude connector directory groups connectors by what they can do. This matters for setting rules: read connectors only pull in context, while write (or read and write) connectors can also take actions in connected systems.
📖Read
- What it does
- Retrieve context, documents or data from connected services.
- Example
- Searching project documentation, pulling meeting notes, referencing knowledge bases.
- Risk level
- Lower — data flows into Claude but no changes are made in the source system.
Read & Write
- What it does
- Retrieve context and also create, update or delete records in connected services.
- Example
- Creating tasks, updating project status, posting messages, writing records.
- Risk level
- Higher — actions change state in external systems. Review and approval controls are especially important.
Interactive
- What it does
- Enable richer interactive experiences within connected workflows.
- Example
- Structured data entry, guided workflows, form interactions in connected tools.
- Risk level
- Depends on what actions the interaction can trigger. Review connector documentation carefully.
Enterprise-Managed Authentication
Check current availability
Anthropic’s documents describe enterprise-managed sign-in as generally available on Team and Enterprise plans. Which identity providers and connectors are supported still varies, so check the setup requirements before rollout.
By default, each person signs in to each connector themselves with OAuth, connecting their own account to the outside service. Enterprise-managed sign-in lets Team and Enterprise admins give people connector access centrally through an identity provider, so staff get access without signing in to each one separately.
Check the current list of supported identity providers and connectors before planning central access. Do not rely on older launch notes that mention Okta only or a waitlist.
Benefits
- • Simplified rollout — employees do not authenticate individually
- • Centralised access management via your identity provider
- • Easier offboarding — remove access from one place
- • More consistent governance across the organization
Considerations
- • Availability depends on supported connectors and identity-provider setup
- • Verify the current identity-provider compatibility list
- • Requires Team or Enterprise plan
- • Check prerequisites and administrator authorization
Connectors, MCP, Projects and RAG: How They Relate
Several related concepts are often discussed together. This table clarifies the distinction to help with both technical and organizational decision-making.
| Concept | What it is | Relation to connectors |
|---|---|---|
| Claude Connectors | User/org-facing integrations that extend Claude with approved external tools, data and actions. | The product experience — what users and admins configure and use. |
| Model Context Protocol (MCP) | The open protocol underlying connector integrations — defines how AI systems communicate with external tools. | The technology that connectors are built on. Not all users need to know MCP; it is primarily a developer/admin concept. |
| Claude Projects | Project spaces that organise conversations, instructions and reference documents; sharing and memory behaviour depend on product settings and plan. | Complementary to connectors. Projects provide stored context; connectors provide live, permission-aware access to external systems. They can be used together. |
| RAG (Retrieval-Augmented Generation) | An AI architecture pattern where relevant documents are retrieved and passed to the model as context at inference time. | Connectors can implement a form of RAG behavior — retrieving context from external systems. RAG as a system-design pattern is broader and can exist independently of Claude connectors. |
| ChatGPT Connectors / Apps | OpenAI's equivalent feature set for connecting ChatGPT to external tools and data. | Separate product with its own terminology and architecture. See the ChatGPT Apps and Connectors guide for comparison. |
Enterprise Use Cases for Claude Connectors
These are common ways organisations use connectors across teams. What is actually possible depends on the connector, your plan and your admin settings.
📋 Read & WriteProject and Task Management
Access project status, create or update tasks, retrieve sprint summaries and surface blockers — without switching between Claude and project management tools.
💬 ReadCommunication and Collaboration
Reference conversation threads, draft responses with context from approved channels, and surface relevant discussions for summarization or decision support.
🎨 Read & WriteDesign and Creative Workflows
Access design files and assets, generate content aligned with brand references, and iterate on creative work with connected design tool context.
🗃️ ReadKnowledge Management
Search internal wikis, documents or knowledge bases so Claude can draw on company knowledge that is relevant to a request, within each user’s permissions.
📊 Read & WriteEngineering and Development
Connect Claude Code to issue trackers, linear boards, databases and internal APIs to support coding, debugging and planning workflows.
🔐 AdminGovernance and Audit
Manage connector access in one place with enterprise-managed sign-in, set connector policies and keep track of which connectors are used and by whom.
What each connector can do depends on the connector, your plan and your admin settings. Check which connectors are available, and their permissions, in your Claude workspace.
Connector Evaluation and Adoption Lifecycle
Organisations that bring in connectors through a clear process, rather than switching them on as they go, tend to keep better control. The steps below are a recommended way to evaluate them.
Evaluation checklist for each connector
- 1 Does this connector serve a documented business need that justifies the data access scope?
- 2 What data classification does the connected system hold — public, internal, confidential or regulated?
- 3 What is the connector's capability type — read only, read & write, or interactive?
- 4 Which user groups need access, and which should be excluded?
- 5 Does the connector respect source-system permission boundaries for each user?
- 6 Who built the connector — pre-built directory or third-party MCP server? Has the third party been assessed?
- 7 Which action policies are appropriate — Always allow, Needs approval or Blocked?
- 8 How will connector activity be monitored and reviewed?
- 9 Does enabling this connector create any compliance implications for regulated data?
- 10 When and how will this connector be reassessed as business needs or risks change?
Security Considerations and Controls
Connecting Claude to outside systems raises security questions your organisation must check for itself. This section covers the common risks and the controls that go with them.
| Risk Area | Recommended Controls |
|---|---|
| Overly broad data access | Apply least-privilege principle — enable connectors only for users who need them; review connected scope. |
| Prompt injection via connected content | Claude has mitigations but is not immune — review retrieved content before trusting it in critical workflows. |
| Unreviewed write actions | Apply Needs approval policy to write connectors; require human confirmation before actions execute. |
| Third-party MCP server security | Anthropic does not audit third-party MCP servers — conduct your own security review before enabling third-party connectors. |
| Credential and token exposure | Use enterprise-managed sign-in where you can, change login keys regularly and watch for access nobody approved. |
| Compliance with regulated data | Enabling a connector does not create compliance — conduct a data flow review for any regulated data touched by the connector. |
| Lack of visibility | Monitor which connectors are active, which users have access and what actions are being taken through connector audit logs. |
Common Mistakes When Deploying Connectors
✗Assuming connectors work the same on all interfaces
Desktop extensions are only available in Claude Desktop and Claude Code — not the web or mobile interface. Before deploying connectors to employees, verify which connector type works on the interface those employees will use.
✗Enabling write connectors without approval policies
Write connectors can create, change or delete records in outside systems. Without a “Needs approval” policy, these actions may run without the user confirming them. Always set review policies on write connectors, especially for live systems.
✗Treating connector responses as authoritative without review
Claude can make errors even with connected context. Connector responses may be incomplete, outdated or misinterpreted. Human review before acting on connector-assisted responses is not optional — it is a governance requirement.
✗Connecting regulated data without a compliance assessment
Switching on a connector does not make you compliant with regulations. If the connected system holds personal, health, financial or other regulated data, you need a formal review of where the data goes and of compliance before rollout.
✗Skipping security review of third-party MCP servers
Anthropic does not check the security of every third-party MCP server. Your organisation must check the security of any third-party connector before connecting it to company data.
✗No offboarding process for connector access
When people leave or change roles, review their connector access and remove it where needed. Enterprise-managed sign-in makes this easier; without it, keep a process for removing each person’s connector logins.
A Practical Connector Adoption Roadmap
Phase 1— Foundation
- → Identify 2–3 high-value, low-risk read connectors for a pilot team.
- → Review data classification of connected systems.
- → Configure connector policies — start with Needs approval for all.
- → Document acceptable use expectations for employees.
— Controlled Rollout
- → Expand read connectors to broader teams after pilot review.
- → Evaluate write connectors with mandatory approval policies and limited users.
- → Conduct employee training on connector use, permissions and responsible usage.
- → Set up monitoring and review of connector activity.
— Scale and Govern
- → Evaluate enterprise-managed auth if on Team/Enterprise plan — centralize provisioning.
- → Review custom connector needs for internal systems not in the directory.
- → Establish periodic connector access review (quarterly or on role change).
- → Assess write connector policies — move from Needs approval to Always allow only where risk is understood and accepted.
Ready to Deploy Claude Connectors at Scale?
Understanding connectors is the first step. Our Claude Enterprise training covers the guide to Claude connectors and permission-aware workflows — permission models, governance frameworks, connector policies and responsible enterprise adoption. For developers, our Claude Code training covers how Claude connectors and MCP integrations work at the engineering level.
Claude Enterprise Training → Claude Code Training →Or explore our Claude AI training course for a structured introduction to Claude across all interfaces.
Frequently Asked Questions
What are Claude connectors? +
Claude connectors are integrations that extend Claude's capabilities by connecting it to approved external tools, data sources and services. Powered by the Model Context Protocol (MCP), connectors allow Claude to retrieve context from external systems, take permitted actions and work within connected workflows — subject to user permissions, administrator controls and the applicable Claude plan.What is the difference between remote connectors and desktop extensions? +
Remote web connectors are cloud-hosted integrations accessible across Claude's web interface, mobile apps, Claude workspaces and Cowork and Claude Desktop and Claude Code where applicable. Desktop extensions run locally on the user's machine and are only available in Claude Desktop and Claude Code — they are not available through the web or mobile interfaces. Organizations should verify which connector type is appropriate for their platform and users.Do Claude connectors give unrestricted access to organizational data? +
No. Claude inherits each user's existing permissions from the connected service — the connector does not grant broader access than the user already has in the source application. Connecting Claude to a system does not bypass that system's own permission controls. Organization owners on Team and Enterprise plans can further restrict what actions connectors are allowed to perform. Claude does not provide unrestricted access to organizational data through connectors.Are Claude connectors available on all plans? +
Connector availability depends on the Claude plan, interface, region, administrator settings and current Anthropic product configuration. Some connectors and connector management features — particularly enterprise-managed authentication and certain organizational controls — are primarily available on Team and Enterprise plans. Connector availability can also depend on whether an administrator has enabled the connector for the organization. Always verify current availability in your specific plan and region.What is enterprise-managed authentication for connectors? +
Enterprise-managed authentication is a centrally managed authorization feature that allows Team and Enterprise administrators to provision connector access centrally — for example, using an identity provider like Okta — rather than requiring each employee to authenticate individually. This simplifies rollout and governance for organizations connecting Claude to multiple systems. Availability, supported connectors and supported identity providers may change; verify current status in your plan's administrator documentation.How does Claude's permission model work with connectors? +
When a user makes a request that involves a connector, Claude operates within the permissions that user has in the connected service. Additionally, organization owners can set policies at the connector level: Always allow (the action can proceed without per-request approval), Needs approval (the user must confirm before the action executes) or Blocked (the action is not permitted). This layered model means the effective permission is the intersection of the user's source-system permissions and the organization's connector policy.What is the difference between Claude connectors and MCP? +
Model Context Protocol (MCP) is the open protocol that provides the underlying technology for connector integrations. Claude connectors are the user- and organization-facing integration experiences built on MCP. Not every user needs to understand MCP to use connectors; MCP is primarily a developer and administrator concept. Custom connectors built by organizations or developers are often implemented as remote MCP servers. See the MCP guide for a deeper technical explanation.Can organizations build custom Claude connectors? +
Yes — organizations and developers can build custom connectors by implementing a remote MCP server, which Claude can then connect to. Custom connectors allow organizations to integrate Claude with internal systems, proprietary data sources or specialized workflows not covered by pre-built connectors. Building custom connectors requires development resources, security review and ongoing maintenance. The specific capabilities and interfaces available for custom connectors depend on the plan and Anthropic's current developer documentation.Does enabling a connector make a workflow compliant with regulations? +
No. Connecting Claude to an external system does not automatically make any workflow compliant with data protection regulations, industry standards or organizational policies. Compliance depends on how data is handled, stored, processed and retained across the full system — including the connected service, Claude's processing and any outputs generated. Organizations should conduct their own compliance assessment for any connected workflow involving regulated data.Does Anthropic security-audit every third-party MCP server? +
No. Anthropic does not security-audit every third-party MCP server or connector. While Anthropic provides documentation and guidelines for building connectors, third-party MCP servers are developed and maintained independently. Organizations should conduct their own security assessment of any third-party connector before connecting it to Claude and organizational data. Enterprise-managed auth and connector allowlists help reduce risk by limiting which connectors employees can access.What is the best way to prepare employees to use Claude connectors responsibly? +
Organizations should train employees to use only administrator-approved connectors, understand which data each connector can access, always review Claude's responses before acting on them, avoid entering sensitive data outside approved connected workflows, recognize that Claude can make errors even with connected context, and follow company acceptable-use policies. Structured Claude enterprise training covering connectors, permissions and governance helps ensure consistent and responsible adoption.Official Claude Resources
These links go to Anthropic’s official documentation. Features and availability may have changed since this was written, so always check the current documents.
Claude Connector Directory ↗
Browse available pre-built connectors from the official directory.
Using Connectors to Extend Claude ↗
Overview of connector types, capabilities and permission model.
Desktop vs Web Connectors ↗
When to use remote connectors vs desktop extensions.
Custom Connectors via Remote MCP ↗
Getting started with building custom connectors.
Enterprise-Managed Authentication ↗
Org-wide MCP connector authorization and current setup requirements.
Claude Code MCP Documentation ↗
MCP configuration for Claude Code (developer-focused).
Claude API MCP Connector documentation ↗
Official documentation for connecting Claude API applications to remote MCP servers and enabling approved MCP tools.
Related Resources
Claude Enterprise Training Claude Code Training Claude AI Training Course What is MCP? ChatGPT Apps and Connectors Guide MCP Training Bangalore Agentic AI Explained AI EngineeringIndependent provider notice: Technovids Consulting Services Pvt. Ltd. is an independent training provider. We are not part of, endorsed by or in any business relationship with Anthropic PBC. Claude, Claude.ai, Claude Code and related product names are trademarks of Anthropic PBC. All product information comes from Anthropic’s public materials. Connectors, features, plans and screens change over time, so check Anthropic’s official documentation for the latest. This page is not legal, privacy or cybersecurity advice.
Connector documentation update
The dedicated enterprise-managed authorization guide describes general availability, while some overview text still uses beta wording. Check the dedicated setup guide and your workspace before deployment. Remote and local access remain different: see Anthropic’s desktop-versus-web connector guide.
Put the guide into practice
Choose where the connection runs
Start with where the tool runs and which version of Claude will use it.
Remote service
Check a hosted MCP server, how it signs you in and what data it sends back.
Take away: A written record of the service and what it is allowed to do.
Review host, client and server rolesLocal tool
Check software running on your own computer, including the files, passwords and programs it can reach.
Take away: An installation and access check that suits the device.
Understand tool capability scope
Match the instructions to your version of Claude
Claude on the web, the desktop app, Claude Code and an app built on the API all connect to tools in different ways. A connection set up in one is not automatically set up in the others.
Anthropic’s guidance on remote MCP says custom remote connections come from its cloud, not from your own computer. That matters when you check network access and private services.
Local does not mean your data stays on your computer
A local connector can read or act on your computer, but what it sends back to Claude may still go to a model in the cloud. Map that whole path before you use sensitive material.
Check who published the software and which features are switched on. Give it the least access that does the job, rather than opening your whole drive for one document task.
Reading, drafting and acting are different
Reading a task list, drafting a status update and posting it are three separate actions. Check which tools really exist, rather than asking for an action and assuming it happened.
Before anything is written or posted, check exactly where it goes and what it says. Think about what happens if it times out: blindly trying again could post it twice.
Worked example: turn task notes into a weekly update
Use a made-up task list with finished work, a blocked item and a missing due date. Ask for an update that keeps those differences and points to the source records.
The missing date should stay unknown. The blocked item should not become “done” because an earlier comment sounded hopeful. Review the draft before any posting.
Try, review and remove connections on purpose
Keep a list of each connection’s owner, publisher, permissions, test users and approved uses. Test what happens when access expires or the connection is removed, not just when it works.
Projects and saved instructions can give Claude context and ways of working, but they do not replace connector sign-in. And working with MCP does not mean a tool is secure. Check access again whenever the task or the team changes.
Check the risk each connection adds
| Situation | Approach | What to check |
|---|---|---|
| Company information held online | Check how remote sign-in works | Identity, permissions and where data goes |
| Files or programs on your computer | Check access on the device | Publisher and the smallest permissions needed |
| Writing to another system | Review before it runs | Target, content and avoiding duplicates |
Practice brief · Fictional data
Create a checklist for approving a connector
Plan the weekly-update exercise with made-up task records, starting with reading only.
What to produce
- A sketch of where data goes and what the connector can do.
- The expected update and the records it points to.
- Checks for disconnecting, expired access and failed actions.
Acceptance checks
- Unknown dates stay unknown.
- Restricted records stay restricted.
- A draft is never described as posted.
- A failure does not trigger an unchecked retry.
This is an illustrative learning exercise, not a client case study or a measured outcome.
Common questions
Is a local connector automatically safer?
No. It can open up things on your computer, and what it sends back may still reach a model in the cloud. Check the permissions and where the data goes.
Does MCP replace business-system permissions?
No. The business system and the connection must still sign the user in and allow only permitted actions.
Continue learning
Turn a pilot into useful team practice
Explore training built around your team’s tasks, approved tools and starting skills. Discuss the scope before selecting a programme.
Explore corporate AI training →Sources and further reading
The examples and checklists here are our own teaching material. Check product-specific steps against the version you use.