An agent can use Capifold to find tools and other agents before you decide which provider to connect. The Capifold MCP connection exposes discovery operations: search for a task, retrieve a candidate and inspect its source information.

This guide uses VS Code's MCP configuration format. Other clients may use different keys. Always start with the configuration shown by the connection guide for your environment; a local development address will not work for a remote user.

1. Copy the actual endpoint

Open the connection guide and copy its client configuration. The public discovery endpoint does not require a Capifold account or API key. A provider returned by the search may have completely separate authentication requirements.

In VS Code, put the configuration in .vscode/mcp.json. The relevant shape is:

{
  "servers": {
    "capifold": {
      "type": "http",
      "url": "COPY_THE_URL_FROM_THE_CONNECTION_GUIDE"
    }
  }
}

Replace the placeholder before using it. Do not paste your provider's API key into this example. VS Code documents setup and available configuration options in its official MCP guide. The project-level file is an example location, not a requirement to commit credentials to a repository.

2. Search for a task or an exact product name

Ask your client to call search_agents. A discovery request could use the following arguments:

{ "q": "document analysis", "limit": 5 }

For a mandatory constraint, make it explicit:

{
  "q": "document analysis",
  "limit": 5,
  "constraints": { "network": "none" }
}

The requirement does not disappear if the search returns no results. Candidates with unknown network behaviour cannot establish “no network access.” Remove or change a requirement only if your actual task allows it.

Read the returned search mode and notice. A lexical fallback compares words when semantic matching is unavailable. It is still a search result, not a successful semantic evaluation. An exact name can find a particular product, but it does not override hard requirements.

3. Retrieve and inspect a candidate

Pass a returned ID to get_agent. Read the original manifest, capability profile, source and version, and open the public profile URL if you want the human-readable view.

Look for the operation you need, supported inputs and outputs, transport, required access and any scoped evidence. Missing facts remain unknown. The evidence guide explains why a provider claim, an inference and a functional check must remain distinct.

4. Prepare the provider connection separately

A profile can link to declared documentation, a provider website or a source repository. If it contains a complete public HTTP MCP endpoint, it can offer a base configuration. Treat that as a starting point: verify required credentials and permissions with the provider before using it.

A package-only offer, another protocol or an endpoint template needs different preparation. Capifold explains the missing step instead of inventing an executable connection. Its discovery operations do not install or run the returned offers.

Troubleshooting without weakening requirements

If the client cannot connect, first compare the configured URL with the connection guide. Use the supported HTTP MCP transport; opening the endpoint as an ordinary web page is not a valid tool call. Check your client's own error message and configuration format.

If search succeeds but no candidate fits, inspect the task wording and requirements. Broaden wording when it changes how you express the same task; do not quietly remove a privacy or access constraint. For a pilot discussion, use the contact link in the connection guide and describe the integration you want to prepare.