How to connect Claude to a company knowledge base

A built-in connector is the smallest operating surface. The employee authenticates to a supported service, Claude uses that person's source permissions, and the organization can restrict connector actions. It is a sensible first route when the business wants cited answers from Google Drive, Gmail, Slack, or another supported system without building a retrieval service.

Enterprise search changes the job from selecting a connector in one conversation to searching across approved organizational sources. Anthropic says the feature can synthesize material from systems such as SharePoint, Slack, Gmail, and Google Drive with source citations. That convenience does not settle source precedence. If a policy page and a signed agreement conflict, the company still needs a rule for which one governs the answer.

Remote MCP and a custom application buy control at the cost of ownership. The team becomes responsible for authentication, tool design, network exposure, logging, failure behavior, and ongoing compatibility. Build that layer for a named gap, not because a custom chat screen looks more serious.

A connector can find a document. It cannot decide that the document is current, approved, or authoritative for the question. Write down the business questions the system will answer and assign an owner, source of truth, freshness expectation, and conflict rule to each material fact.

The distinction matters when the same customer term appears in a proposal, CRM note, Slack thread, and signed contract. Search relevance may favor the most explicit passage while the business rule favors the executed agreement. Retrieval quality and authority are different decisions.

If the team cannot name the governing record, start with source repair. A fluent answer over contradictory material can increase confidence without reducing ambiguity.

Anthropic's current connector documentation says connectors inherit each person's permissions from the connected service. Team and Enterprise owners can also allow, require approval for, or block connector actions. Those are useful controls, but they need an end-to-end test.

Test a normal employee, a person in another department or client account, a guest, and a recently revoked user. Inspect the retrieved source and citation, not only the final sentence. A correct-looking answer can still prove that the wrong identity or shared credential reached the record.

For a custom application, record where the user identity becomes a source-system filter. A broad service account can work only if the application applies the employee's narrower boundary before retrieval and keeps evidence of that decision.

A demonstration normally asks a question whose answer is present and permitted. An acceptance test needs the opposite cases. Ask for a record the user cannot access. Remove a permission and repeat the query. Present two approved sources that disagree. Ask for a fact that does not exist. Place an instruction inside a retrieved document that asks the model to reveal another source or call a tool.

The expected result should be written before the run. A missing fact should produce a limitation. A conflict should expose both records or follow the documented precedence rule. Retrieved text should remain evidence, not become authority to change tool access.

• Current approved source with a known answer
• Restricted source requested by an unauthorized role
• Revoked user or permission changed after connection
• Two approved records with conflicting terms
• Missing evidence that should stop the answer
• Retrieved content containing a hostile instruction
• Write request whose approver is unavailable
• Destination failure after approval

Search and citation create a bounded first release. A write action creates a different failure surface. Sending an email, changing a CRM field, or creating a calendar event needs a permitted purpose, structured input, validation, approval where consequence warrants it, idempotency, and a record of the destination result.

Anthropic lets Team and Enterprise owners narrow connector tool permissions. Use that boundary. A connector that can read a policy does not need permission to create documents, and a sales workflow that can draft a CRM update does not need unrestricted account administration.

Only add a write after the read path passes its test set. The action should be approved and evaluated on its own cases rather than smuggled into a successful search pilot.

For many companies, the right answer is a managed Claude workspace with two or three approved connectors and a documented source map. That route is faster to reverse than a custom platform and exposes whether the underlying permissions and records are ready.

Move to remote MCP when a trusted unsupported service or a narrow tool blocks the useful workflow. Move to a custom application when the company needs its own interface, tenant-aware retrieval, specific evidence ranking, evaluation, or a controlled action route. The custom layer should purchase one clear operating capability.

If source ownership and retrieval need to be designed before Claude is connected, start with the broader company knowledge-base implementation guide. Connection is a delivery choice inside that larger system.