Remote MCP servers: local vs hosted, and how to connect

Remote MCP servers expose tools from a network endpoint instead of a process launched beside the AI client. They make hosted services reachable from compatible clients, while local servers are useful for files and private environments. AIPx provides a hosted, OAuth-protected endpoint for supported service accounts and existing public custom MCP connections.

What is a remote MCP server

Remote describes where a server is reached. Transport describes how messages move. A remote server commonly uses Streamable HTTP, while a local process commonly uses stdio. A server running on your machine can also listen on localhost HTTP, so the terms are related without being identical.

The client discovers the endpoint’s capabilities and tools, then sends calls with the required authentication. The server executes the operation or forwards it to a service it wraps. The returned data becomes part of the agent’s task context; its origin being remote does not make it more or less trustworthy by itself.

For example, a coding agent can retrieve an issue from a hosted tracker without starting a tracker server on the laptop. A local filesystem server, by contrast, can be launched with access to one directory on the laptop. They solve different reachability problems.

For the protocol role rather than placement, see what an MCP server is. The official local-server guide describes the local process route.

Local vs remote MCP servers

Decision	Local server	Remote server
Runtime location	A process in your environment	A reachable network service
Common transport	stdio	Streamable HTTP
Installation	A command, runtime and package or binary	A URL and authentication configuration
Private files	Can access the paths you grant	Needs an explicit way to reach the data
Several clients	Each client needs compatible access or launch configuration	Clients can reach a shared endpoint
Operational burden	You maintain the local runtime	A service operator maintains the endpoint
Main dependencies	Local executable, environment and data access	Network, hosted service and upstream availability
Trust boundary	Code running in your environment	The remote operator and connected providers

Local is a sensible choice when a tool needs a private repository, folder or internal service reachable only from your environment. Remote is convenient when the data already lives in online services and you want several clients to use a connection. Neither choice eliminates the need for narrow credentials and useful tool descriptions.

Localhost is relative to the caller

localhost means the machine making the request. If a hosted client tries to reach your laptop’s localhost address, it reaches its own environment instead. Merely opening a desktop interface does not establish a network route between that cloud client and your computer.

A tunnel or private network route can change reachability, but it also creates a new access boundary. Do not expose an unauthenticated development server simply to make a menu’s URL test pass. Choose the deployment and authentication together, especially for tools that can execute commands or modify data.

MCP hosting options

Hosting ranges from a vendor-managed endpoint to a custom server you operate. Choose the smallest arrangement that meets the data and permission requirements.

Option	You operate	Useful when	Tradeoff
Official hosted server	Account configuration	One vendor already provides the needed tools	Its account model and tool inventory define the boundary
Local server	Runtime and package updates	Tools need files or private local systems	Setup repeats across environments
Self-hosted remote server	Deployment, authentication and maintenance	A custom tool needs shared network reach	You own availability and security work
Local container gateway	Container runtime and configuration	You want managed local process isolation	It still depends on your local environment
Hosted gateway such as AIPx	Account and agent rules	Several hosted accounts need scoped access	Another operator and network dependency

Docker’s MCP Gateway documentation describes a local container-based approach. It is a useful alternative when running and isolating local servers is part of your requirement. A hosted account gateway is a different operating model, rather than a direct replacement for every local runtime feature.

Self-hosting can fit an internal API whose data must stay within a controlled environment. It also means maintaining credentials, validating tokens, monitoring failures and updating the server. A reachable HTTPS URL is only the beginning; a deployed server still needs an authorization model and a tool surface that fits the agent.

A direct official endpoint can be the simpler choice when you need one service and one account. A gateway starts to help when several accounts or clients need different policy views. The MCP gateway guide compares that connection layer with servers, proxies and routers.

Claude remote MCP

Claude’s remote connector requests originate in Anthropic’s cloud, including when you use the desktop app. Use a public reachable endpoint for the remote connector path. Local desktop process configuration is a separate route and does not automatically carry over to the web application.

To connect AIPx:

1. Connect the service account in AIPx.
2. Create an agent and review its account and tool rules.
3. Copy https://dash.aipx.run/mcp/AGENT_ID with the dashboard’s actual identifier.
4. Add it through Customize → Connectors → Add custom connector.
5. Complete OAuth and try a verifiable read.

Anthropic documents its remote connector workflow. The Claude setup guide covers the current interface and account considerations. For terminal work, use Claude Code’s HTTP setup; the connector menu and terminal registration are different configuration paths.

For Codex and Cursor, remote connection uses their own file and command formats. Follow Codex MCP setup and Cursor MCP setup instead of pasting a desktop local-server entry into either client.

How AIPx hosts MCP for your accounts

AIPx provides the client-facing agent endpoint and connects its allowed tools to supported service integrations. You connect accounts once, then give agents different views over them. The endpoint is protected by client OAuth, while upstream services retain their own account authorization requirements.

The URL identifies an agent rather than acting as a secret API key. A client still needs to complete authorization. Two similarly named endpoints may represent different permission sets, so copy the correct one from the dashboard instead of reconstructing it from memory. The MCP OAuth guide explains those authentication boundaries.

AIPx also accepts a public URL for an existing custom MCP server. Automatic connection can discover supported OAuth registration or no-auth access, and the product can use a bearer token where configured. The upstream server must already be running and reachable.

This is not a hosting service for your arbitrary stdio process, source repository or Docker image. If you have a private local server, operate it in an appropriate environment first. If you have a REST API without an MCP interface, the API-to-MCP guide explains the separate conversion step and the limits of AIPx’s supported routes.

Permissions on a hosted connection

Rules combine tool, account and agent. Hide is enforced server-side and removes the operation from that agent’s surface. Allow and Confirm are execution hints; Confirm asks the client to obtain approval, and actual review depends on the client’s behavior. Use Hide for an operation that must be unavailable rather than relying on an unverified approval workflow.

A recommended starting set is a narrow group of relevant reads with writes marked Confirm in a client that honors it. Check confidentiality as well: reading a mailbox or health account can reveal private information without modifying the service. The permission guide explains that distinction.

The audit records which agent, account and tool ran, with parameters, result and time. Inspect a first call to verify the connection used the expected source. Do not infer retention, exports or availability guarantees from the existence of that audit.

What costs remain separate

AIPx Free includes three connected accounts and one agent. Pro costs $3 per month for unlimited accounts and agents plus Pro integrations. Upstream subscriptions, API usage restrictions and rate limits remain the service provider’s concern. A hosted endpoint does not turn every connected API into unlimited free access.

The integration catalog lists supported services. Begin with those whose data is already needed for the task, such as a tracker and an error system for debugging, or mail and calendar for meeting preparation.

Troubleshooting remote connections

When a connection fails, check reachability first: the exact HTTPS endpoint must be accessible from the client’s actual runtime. Then inspect client OAuth, discovered tools and agent rules, followed by the upstream service connection. A provider denial after login is not evidence that the public MCP URL is wrong.

A connection with no expected tools can be a policy or discovery issue. Refresh after a configuration change, and inspect Hide rules before allowing extra operations. A transient network failure should not prompt you to replace OAuth with an exposed token.

Use a local or privately operated server when data placement requires it. Use hosted account tools when avoiding local runtime management is useful and the operator trust boundary is acceptable. Connect free.