Claude Connectors and MCP consulting

A connector is the link between Claude and an outside system. Under the hood almost every connector is an MCP server: a service that lists the tools it offers, such as search a mailbox or look up an order, and runs them when Claude asks. Claude reaches company systems in three ways. You can add a listed connector from the Connectors directory, add any remote MCP server by its URL as a custom connector, or have a server built for a system that has no connector at all. On Team and Enterprise plans an Owner adds each connector once for the organization, and every member then signs in with their own account.

1. 01

   Directory connector, custom connector, or a server we build

   Start with the directory. Microsoft 365, Google Drive, Gmail, Google Calendar, Slack, Atlassian, Notion, Asana, Box and HubSpot all have listings, and the vendor maintains the server. Use a custom connector when a vendor publishes an MCP endpoint that is not listed, or when you need your own OAuth client in front of it. Build a server when the system is yours: an ERP, a SQL warehouse, a ticketing queue, a pricing engine. That server must be reachable from the egress IP range Anthropic publishes, because Claude connects from the Anthropic cloud, not from the employee's laptop. If it lives inside your network, Enterprise organizations can request access to MCP tunnels, which open an outbound-only connection instead of a public endpoint.

2. 02

   Where connector projects go wrong

   The common failures are predictable. Someone adds a custom connector with an API key in a request header, and every member now acts with that key's access instead of their own. Write tools are switched on during the pilot because they demo well, and nobody reviews which ones Claude may call without asking. Research mode calls connector tools automatically without further approval, so a write-capable tool left enabled can fire during a long research run. The Microsoft 365 connector quietly fails a new feature because the tenant consented before Anthropic added a permission. And the internal server works on the VPN but not in Claude, because traffic from the Anthropic IP range cannot reach it. Each of these is avoidable with a written access design before the first connection.

3. 03

   Why Abstrakt Solutions

   We have built Salesforce integrations since 2017, and connector work is integration work with a model on one end. We start with an inventory of systems, owners and data classes, decide per system whether Claude reads, drafts or writes, and test the OAuth flow with your identity team before anyone outside IT sees it. We build custom servers with the official MCP SDKs, put OAuth in front of them rather than shared keys, and log every tool call on the server side. Salesforce-specific MCP work is covered separately on our MCP and Salesforce guide; this service covers everything else Claude needs to reach. Abstrakt is a Certified Services Partner in the Claude Partner Network with 12 Claude certifications on the team.

The engagements we run most often on Claude Connectors and MCP, from first implementation through optimization.

1. 01

   Inventory systems and decide the access level

   We meet with IT, security and two or three department leads for a working session. For each requested system we record the owner, the data classes it holds, the identity provider in front of it, and whether Claude should read, draft for human review, or write. Anything holding regulated data gets a named approver. The output is an access matrix and a short list of first connectors, usually one productivity suite and one internal system.

2. 02

   Set organization controls before anyone connects

   An Owner adds only the approved connectors in Organization settings > Connectors. For Microsoft 365, your Entra Global Administrator grants tenant consent and we confirm write actions stay off for the pilot. We set tool permissions so anything that sends, deletes or updates needs approval or is blocked. We also turn off write-capable tools for research use, because research mode calls connector tools without asking.

3. 03

   Run a read-first pilot with a small group

   Ten to twenty people from two departments use the connectors on real work for two to four weeks. They log wrong answers, missing documents and permission surprises in a shared tracker. We review weekly with IT. Most issues at this stage are search quality, such as unindexed SharePoint sites, or permission gaps in the source system that were invisible until Claude found the files.

4. 04

   Build and secure custom servers

   For internal systems we build a remote MCP server with a small set of narrowly named tools, OAuth through your identity provider, and token audience checks as the MCP authorization spec requires. We host it where the Anthropic egress range can reach it, or behind an MCP tunnel. Each tool call is logged with user, tool, arguments and result status, and the logs go to your SIEM before the server opens to the pilot group.

5. 05

   Turn on writes deliberately and hand over

   Once read accuracy is acceptable, we enable specific write tools for specific groups, starting with drafts rather than sends. We document how to add, remove and re-authorize connectors, what to do when a vendor changes its endpoint, and who reviews new connector requests. Your admins run the next connector themselves with our runbook, and we stay available at our hourly rate for anything new.

1. 01

   Sharing one API key across the company

   A request-header credential gives every member the same access. Use OAuth so each person acts as themselves, and keep shared keys for low-risk, read-only services.

2. 02

   Leaving write tools on during the pilot

   Send, delete and update tools make good demos and bad surprises. Block or require approval for them until read accuracy has been reviewed.

3. 03

   Clicking Always allow on untrusted servers

   Anthropic warns that malicious servers can carry hidden instructions. Reserve Always allow for tools you would let run unsupervised, and watch for tool changes after a server updates.

4. 04

   Assuming the VPN route will work

   Claude connects from the Anthropic cloud, not the user's device. Plan the network path, egress allowlist or tunnel, before you build the server.

Other Claude platforms we implement

• Is MCP safe for business use? The real risks and the controls that work
• MCP and Salesforce explained: connecting AI assistants to your CRM
• How do you connect Claude to your company's data? SharePoint, Google Drive, Slack, databases and ERPs