Docs · Updated 2026-10-03

Local vs remote MCP servers: which should you choose?

Choose a local MCP server when the task needs access inside your client’s execution environment. Choose a hosted remote server when a reachable service already provides the tools you need. Check the client, data access and permissions before connecting.

A directory listing helps you find a server. It does not determine where that server runs or give your agent permission to call it. ToolCargo offers both MCP discovery and its own hosted connectors; those are separate connections.

What do local, stdio and Streamable HTTP mean?

With stdio, the MCP client launches a server process and exchanges messages through its standard input and output. That process runs in the client’s execution environment, which could be your laptop, a container or another machine running the client. See the official stdio specification.

With Streamable HTTP, the client sends MCP messages to a server endpoint over HTTP. A hosted service normally supplies an HTTPS URL. HTTP can also serve a process on localhost; the transport alone does not tell you who operates it or what data it can access. See the official Streamable HTTP specification.

Local vs hosted remote: a decision table

Typical setups; inspect the particular server and client.
Question	Local server over stdio	Hosted remote server
Where does it run?	Where your client launches the process.	In the service operator’s environment.
What do you configure?	A command, runtime and any required environment variables.	A server URL and its required authentication.
What data is reachable?	Resources allowed by the process, OS and server configuration.	Resources the service can reach and is authorized to use.
Who maintains the server?	You manage the local package and runtime.	The operator maintains the hosted service; you manage your connection.
What must the client support?	Launching that server and communicating over stdio.	The endpoint’s transport and authentication method.

Local execution does not guarantee that data stays on your computer: a local server can call external APIs, and your AI application has its own data handling. A remote URL does not grant access to your laptop’s files. Review the actual server’s behavior and access boundaries.

Choose from the task you need to complete

Read a folder on your laptop: choose an appropriate local filesystem integration and explicitly configure its allowed directories. ToolCargo’s hosted endpoint cannot read a path on your laptop. The official local-server guide explains the filesystem example and process permissions.
Audit a public website: ToolCargo Site Audit runs remotely against public URLs. Its documented checks and limits describe what it reads.
Read an account in a cloud service: check whether the chosen connector requires a provider connection, which operations it offers and which permissions it requests. The URL alone is insufficient.

Find a server without confusing discovery with execution

In the ToolCargo directory, select Remote URL to find entries with remote endpoint metadata, or Local package to find entries with package metadata. A listing can offer both. These filters reflect a dated registry snapshot, not a live health check or a guarantee that the server works in your client. Inspect the publisher’s setup details.

ToolCargo’s public discovery MCP needs no account or key:

Discovery server URL

https://toolcargo.com/mcp

It searches listings and returns connection details. To run a listed server’s tools, connect that server separately. External providers set their own permissions, credentials and costs.

ToolCargo hosted tools have their own connector endpoints. For example:

Site Audit hosted server URL

https://toolcargo.com/mcp/site-audit

That endpoint requires ToolCargo authentication and uses your plan’s allowance. Follow getting started and the client setup and test-status page. Check the selected connector’s requirements before granting access.

Keep a short connection decision

Decision template

Task:
Data the task needs and where it lives:
Chosen server and publisher:
Local command or remote endpoint:
Client transport and authentication support:
Allowed operations and permissions:
Provider costs and ToolCargo quota, if applicable:
One small test and its observed result:

Start with a small operation whose result you can check. Finding a listing or adding a URL does not prove an end-to-end tool call succeeded. Keep client support and the exact server’s test status separate.

Common questions

Can a local-only client connect to a remote MCP server?

A local bridge can forward requests to a remote endpoint. It adds a local process; the selected remote tools still execute on the remote service. ToolCargo’s bridge example documents this option and its test limits.

Is a hosted MCP server automatically compatible with every client?

No. Client transport, authentication and protocol-version support still matter. The official remote-server guide explains connection and permission review; use your client’s current instructions for its exact setup.

Does ToolCargo host every server in its directory?

No. The directory includes external publisher listings. The hosted connector catalogue identifies ToolCargo’s own tools and distinguishes available products from planned ones.