Local vs remote MCP: where does the work actually run?
A local MCP server runs beside its client; a remote MCP server is reached over a network transport. Neither description, by itself, tells you which computer’s files a tool accesses or what a command may change. Identify the execution host, authorization and data path before choosing a connection.
By Ganado · Published 20 September 2026 · Updated 22 September 2026 · Product scope: Ganado Bridge 0.2.0-preview.0, including its optional hosted technical preview. Public directory status and supported client behavior are separate from the protocol.
The decision in one table
| Question | Local stdio process | Remote service |
|---|---|---|
| How does the client reach it? | It starts a process and exchanges protocol messages through its input/output streams. | It connects to a service endpoint, commonly using Streamable HTTP. |
| Which files can it reach? | Whatever the process’s code and actual OS access permit on its execution host. | Whatever the service is implemented and authorized to reach; not automatically the laptop in front of you. |
| Does it need a public URL? | Not for the direct local stdio connection. | A network address reachable by the intended client, with appropriate authentication. |
| Is it a sandbox? | Only if a separate, effective boundary is implemented. Local does not mean restricted. | Only if service isolation is actually implemented and verified. Remote does not mean isolated. |
The protocol’s host/client/server roles and the stdio/Streamable HTTP transports are documented in the MCP architecture overview. The consequences in the remaining sections are practical checks derived from those distinctions, not universal promises about every server.
A cloud assistant does not automatically have your Mac
Consider a prompt asking an assistant to edit /tmp/example.txt. The pathname is incomplete operational context: it identifies a file relative to a particular machine’s filesystem. If the tool runs on a cloud host, a similarly named path there is not your local file. A tunnel or device agent can bridge that gap, but it introduces a separate connection and authorization step.
Before changing anything, ask the tool to identify the execution host. Then name a disposable file on that host, read it, make one checked change and read it back. This is a better first acceptance test than assuming that a familiar hostname or a connected badge proves full access.
What the downloadable Ganado Bridge actually supplies
Bridge 0.2 includes a stdio agent for an authorized local MCP client and an optional client for the hosted technical preview. Both routes use the same file tools, bounded search, checked edits and command sessions on your paired computer. The hosted route uses OAuth and a short-lived pairing code. It does not provide a hosted computer, guaranteed uptime or an approved ChatGPT directory listing.
Its direct file tools operate on the connected host. Remote shell execution is possible only for existing SSH aliases that the owner explicitly configures; none are enabled by default. That optional SSH operation does not silently turn local file tools into remote file tools. The technical overview and release package specify the current behavior.
“Local” does not mean that no data leaves the computer
The file may remain on the Mac until a tool reads it. The tool response can then contain its contents, path or image bytes and be passed to the AI provider chosen by the client. That is a data transfer even when the server itself is local. Decide whether the requested material is appropriate for that provider before asking for it.
Bridge’s security model explicitly describes owner-level access and the fact that shell commands may change or delete files. Its privacy notice separates website measurement from agent execution. The local agent does not add automatic usage telemetry, but your AI provider and host have their own handling rules.
A directory entry is discovery, not a working connection
A package appearing in the MCP Registry means that its metadata is published there. It does not establish approval by every AI platform or install that package into a user’s account. Bridge’s local package is published in the MCP Registry; it is not currently published in the ChatGPT directory.
OpenAI has a separate submission and review process with requirements for the connection and reviewer access. See the official submission requirements. The local agent is free under MIT; no paid licence is required. The hosted technical preview and a future directory approval remain separate facts.
Choose the smallest connection that completes the task
For a technical user evaluating work on one Mac, start with the local package and a client that supports stdio. Verify a file change and a completed command before involving remote servers. The Cursor setup guide shows one configuration route, and the troubleshooting guide distinguishes launch, protocol and tool failures.
For access from ChatGPT, use the Developer Mode setup guide if that feature is available on your account. Extract the bundle, run node bin/bridge.mjs connect, add the MCP URL with OAuth and authorize the displayed code in the browser. Keep the Mac awake and the connection running. The MIT licence includes commercial use, and the hosted preview currently requires no payment.