Anthropic adds remote MCP support to Claude Code

Anthropic announced remote Model Context Protocol support in Claude Code. The company said developers could connect Claude Code to remote MCP servers by adding a vendor URL instead of operating local servers.

By Hypler editorial team · Source: Anthropic.

Original source date: .

Hypler publication: . Hypler updated: .

Estimated reading time: 4 minutes.

Topics: Anthropic, MCP, Claude Code, OAuth

Anthropic logo

Generated editorial illustration: a developer workspace connected to a separate locked tool workspace.
Generated editorial illustration: a developer workspace connected to a separate locked tool workspace. Credit: Hypler / AI-generated editorial illustration. Hypler editorial asset

The remote server remains a separate service

Anthropic announced remote Model Context Protocol support in Claude Code on June 18, 2025. It said developers could add a vendor URL instead of operating a local server, then use exposed tools and resources. The announcement gave examples involving development tools, project-management systems, and knowledge bases, including Sentry and Linear, and described native OAuth. It placed updates, scaling, and availability with the remote-server vendor. The connection may remove local maintenance, but it introduces a service with its own identity model, capability catalog, change history, and failure behavior.

Treat the connection as a capability contract

MCP makes operations discoverable; it does not decide which should be permitted. Record each tool's input schema, output type, data classification, action effect, owner, and approval requirement. Read-only issue lookup carries different stakes from changing a ticket, editing a record, or publishing a release. OAuth establishes an authenticated connection, while the application still chooses scopes, session lifetime, revocation, and how a request maps to an identity. The endpoint, configuration version, approved scopes, and actual enabled tools should remain visible in an operational record, ready for review when the server changes.

Discovery must not broaden authority silently

Store remote-server configuration as reviewed application state: endpoint, transport, allowed tools and resources, authentication method, scope set, timeout, and owner. Resolve metadata before use and compare the discovered catalog with the approved allowlist. A new tool, changed schema, or broader resource pattern should trigger a review rather than silently enlarge the agent's surface. For a consequential call, the model can prepare a structured request containing target, parameters, evidence, and reason. A policy layer still validates that request against both the permitted operation and the caller's authority before the receiving system acts.

Read-only issue context is a useful first boundary

Imagine a hypothetical code-review workflow that receives a pull-request identifier and can call one read-only issue-tracker tool. It fetches the linked issue title, acceptance checks, and approved labels, then includes those fields beside the source diff in a draft review. The output cites the issue identifier and revision. The connection has no tool for changing issue state or adding comments. An access error or repository mismatch becomes a recorded gap; a reviewer may correct the link or proceed with source-only review. The external system contributes context without becoming an excuse to expand access.

The endpoint is an input boundary, not policy

A remote server can alter its catalog, return incomplete resources, redirect a request, or go down. Returned text can contain imperative language that a model mistakes for trusted direction. Treat it as data, validate tool arguments against schemas, and apply action policy after a response as well as before it. Test expired authorization, revoked access, added tools, changed schemas, malformed responses, delays, and a resource asking the agent to do unrelated work. The desired result is a clear error category tied to the original task, with state-changing work still pending review.

Integration evidence should outlast setup speed

Build fixtures for expected resources and responses, including permission failures and changed fields. Confirm that the agent calls only allowed operations, makes use of returned information traceable, and remains in scope when the external system fails. Review OAuth scopes, account ownership, logs, and revocation with the connected service's owners. A rollout can start with a read-only server in a limited workspace and telemetry for connection failures and tool calls. Consequential tools need separate tests for receiving-system behavior, authorization, retries, and receipts before they are enabled.

Convenience does not reduce accountability

A vendor URL may spare the team from hosting a local server; it cannot spare the team from knowing who may use which remote effect. Hypler's analysis treats remote MCP as a reviewed software dependency with bounded capabilities and accountable ownership. The historical announcement does not establish that a specific server is available, safe, or authorized for a specific task. Revisit the connection when its catalog, OAuth configuration, or workflow changes. That is where the actual reliability and security question resides.

Hypler analysis

Hypler’s view

Remote context is remote capability. Adding a URL can add identities, scopes, data paths, and side effects to an agent, so each server connection deserves a dependency review rather than a convenience exception.

Hypler technical analysis

Engineering notes

Remote MCP moves tool discovery and operation from a local process into a networked provider boundary. A vendor URL and OAuth connection therefore become part of the system's authorization graph: client identity, granted scope, exposed operations, returned data, and action evidence require independent inspection.

What to inspect

Server capability inventory
List every exposed tool and resource, its input schema, effect type, data domain, and whether it can read, write, or trigger an external action.
OAuth grant
Inspect client identity, consented scopes, token storage path, renewal behavior, revocation path, and the resource account represented by the connection.
Call evidence
Trace a denied request, a changed server response, and a retried action; inspect logs for selected operation, supplied arguments, authorization decision, and resulting outcome.

Scope and limits. Remote support lowers local server operation. It does not verify vendor service behavior, make a tool appropriate for a task, or establish permission to use connected data.

Terms and definitions

MCP
Model Context Protocol, an open protocol for connecting AI applications to tools, data sources, and prompts through defined interfaces.
Tool Calling
A model interface pattern where the model requests a structured operation from an external tool.
API Layer
The server-side interface that exposes defined data and actions to clients or integrations.
Testing and Observability
Testing checks expected behavior; observability uses runtime signals to understand a system’s actual behavior and condition.

More on this subject

Generated illustration of isolated tool-execution fixtures and a controlled test gate; not deployed software.
Generated illustration of isolated tool-execution fixtures and a controlled test gate; not deployed software. Credit: Hypler / AI-generated editorial illustration. Hypler editorial asset; not vendor or event photography

Original source

Historical source briefing with Hypler engineering analysis. Source details reflect the original publication date; implementation examples are hypothetical. Technical analysis expanded October 24, 2026; this does not change the original source event or establish a new source verification. Editorial prose revised October 26, 2026; original event, publication and source-check dates are retained.

Related capabilities