AI agents do not find MCP servers on their own. A person or developer configures each server for the agent, by URL or in a config file, and public MCP registries are where those servers are published so humans can discover them. As of mid-2026, no mainstream agent roams the internet auto-connecting to tools. Discovery and connection are two separate steps, and both put a human in the loop. Here is how each one actually works.
Do agents auto-discover servers?
No, and that is by design. An agent like Claude or ChatGPT can only call tools from servers that have already been configured for it. Someone adds the server, once, and from then on the agent sees its tools in every conversation. The agent is not scanning the web, guessing URLs, or negotiating access with strangers. This is a safety boundary as much as a technical one: letting a model connect itself to arbitrary software as you would be reckless. So the model decides which of its available tools to call; a human decides which servers are available in the first place. If you are new to the underlying idea, our explainer on what an MCP server is covers the plumbing.
How does a server actually get configured?
There are two shapes, depending on where the server runs. A remote (hosted) server is added by URL: you paste its address into the client and sign in. LedgerMCP is this kind, added at https://ledgermcp.com/mcp with a 6-digit email code or an lmcp_ API key. A local server runs as a subprocess on your own machine and is added through a config file that names the command to launch it. Either way, a human does the wiring:
- Remote server. Add the URL in the client, authorize with OAuth or a bearer key. Nothing to install. This is the fast path for non-developers.
- Local server. Install the server, then register the launch command and any credentials in the client's config file. More setup, more control, one machine at a time.
This is what people mean by runtime configuration: the concrete connection an agent uses at the moment it runs. For the step-by-step version in Claude, see how to connect Claude to an accounting MCP server.
What are MCP registries, and who runs them?
Registries are directories where server authors publish their servers so people can find them. As of mid-2026 the main ones are:
| Registry | What it is |
|---|---|
| Official MCP Registry | registry.modelcontextprotocol.io, the central machine-readable catalog (preview since September 2025) |
| Smithery | A large browsable directory that indexes and ranks servers for discovery |
| Glama | A crawler-backed directory tracking tens of thousands of servers |
| PulseMCP | A curated directory and one of the ecosystem's trusted contributors |
| GitHub MCP Registry | A sub-registry surfacing servers published to the official registry inside GitHub |
As of mid-2026 the official registry is meant to be the source of truth, with the others acting as human-friendly front doors and sub-registries that build on top of it. Publish once to the official registry and compatible directories can pick the entry up.
Registry discovery versus runtime config: what is the difference?
These two are easy to conflate, so keep them apart. A registry helps a person find a server: you browse, read the tool list, and decide it is worth trying. Runtime configuration is how the agent connects: the URL and credentials you enter in the client. A registry listing is a business-card; the config is the actual phone line. Finding a server in a directory does nothing until you configure it, and you can configure a server that is in no directory at all (a private internal server, for instance). As of mid-2026, the agent itself does not silently query a registry mid-conversation to add new powers; a human moves a server from "listed somewhere" to "connected here."
How does this work for LedgerMCP?
LedgerMCP is a hosted remote MCP server that you add by URL. You do not wait for an agent to discover it; you point your agent at https://ledgermcp.com/mcp once, sign in, and it is available from then on. That gives you 60 bookkeeping tools (24 read, 36 write) over the same connection every time, with the guardrails enforced in the database rather than in the model. Human documentation lives at the docs, and machine-readable instructions for agents live at /llms.txt, generated from the same tool registry so the two never drift. The accounting MCP server page has the full tool breakdown. The takeaway holds across every client: discovery is a human activity, connection is a human decision, and the agent works only with what you have deliberately handed it.
Quick answers
Do AI agents discover MCP servers on their own?
No. As of mid-2026 an agent only uses the servers a person or developer has configured for it, by URL or in a config file. It does not roam the internet finding and connecting to tools by itself. Discovery of new servers happens through registries and directories that people browse, not through the agent.
What is the official MCP Registry?
The official MCP Registry (registry.modelcontextprotocol.io) is a central, machine-readable catalog of publicly available MCP servers, launched in preview in September 2025. As of mid-2026 it acts as the source of truth that downstream directories and clients build on, rather than a place your agent silently queries.
What is the difference between runtime config and a registry?
Runtime config is what actually lets an agent use a server: the URL and credentials you enter in the client. A registry is a directory where servers are published so people can find them. The registry helps you discover a server; the config is how your agent connects to it.
How does an agent connect to LedgerMCP?
You add it yourself. LedgerMCP is a hosted remote MCP server you connect by URL at https://ledgermcp.com/mcp, using OAuth sign-in in Claude or ChatGPT, or an lmcp_ API key in other clients. The agent does not find it on its own; you point the agent at it once and it stays available.
