Annotate publisher plugins with Copilot connector pointers

[dhruvchand]

Summary

Lets clients (the Copilot app) find a tool such as Figma through Agent Finder and learn that a managed Copilot connector exists for it, so they can prefer the connector.

The catalog stores only a static pointer to the connector's canonical name. Availability depends on the user and tenant and is feature-flagged, so clients check eligibility at runtime with the connector service (GET /plugins). The catalog makes no availability claims.

Shape

Each connector-backed entry gets a copilot-connector tag plus metadata.copilotConnector:

"tags": ["copilot-connector"],
"metadata": {
  "sourceSet": "figma/mcp-server-guide",
  "repoPath": ".github/plugin/plugin.json",
  "copilotConnector": { "name": "githubcopilotfigmamcp" }
}

• metadata is used because Agent Finder's CatalogEntry wire model drops unknown top-level fields. The earlier top-level connector and matchHints fields never reached clients.
• There is no available flag, no matchHints, and no top-level connector field.
• scripts/generate_ai_catalog.py validate() now checks that metadata.copilotConnector is an object whose only key is name, matching ^[a-z0-9]+$. An entry must have both the tag and the metadata, or neither. Unit tests cover a valid entry, a missing name, a bad name, an extra key, a tag without metadata, and metadata without a tag.
• The convention is documented in CONTRIBUTING.md.

Connector coverage (12/12)

Most publishers ship an official plugin that bundles the same hosted MCP server the connector proxies, so this PR annotates those plugin entries.

Connector name	Catalog entries	Hosted MCP server
githubcopilotfigmamcp	catalog/figma/figma.json (new)	mcp.figma.com
githubcopilotatlassianrovomcp	catalog/atlassian/atlassian.json (new)	mcp.atlassian.com
githubcopilotcloudflaremcp	catalog/cloudflare/cloudflare-plugin.json (new)	mcp.cloudflare.com
githubcopilotlinearmcp	catalog/linear/linear.json (new), catalog/anthropics/linear.json	mcp.linear.app
githubcopilotnotionmcp	catalog/makenotion/notion.json (new)	mcp.notion.com
githubcopilotsentrymcp	catalog/getsentry/sentry.json (new)	mcp.sentry.dev
githubcopilotslackmcp	catalog/slackapi/slack.json (new), catalog/anthropics/slack-by-salesforce.json	mcp.slack.com
githubcopilotvercelmcp	catalog/vercel/vercel.json (new)	mcp.vercel.com
zapiermcp	catalog/zapier/zapier.json (new)	mcp.zapier.com
microsoftlearndocsmcpserver	catalog/microsoftdocs/microsoft-docs.json (new)	learn.microsoft.com/api/mcp
workiqmcp	catalog/microsoft/workiq.json	—
kusto	catalog/github/kusto.json (new, data-backed; no publisher plugin exists)	—

For each new plugin entry, I checked the default branch (main), the manifest path, and the bundled .mcp.json/mcp.json against the upstream repos. Descriptions come from the publisher manifests. The Cloudflare identifier ends in :cloudflare-plugin because urn:ai:github.com:cloudflare:skills:cloudflare is already used by an existing skill entry.

Why the connector-annotations.json overlay was removed

• It targeted urn:air:registry.modelcontextprotocol.io:* rows. The live Agent Finder serves MCP entries as urn:air:api.mcp.github.com:* from a different source, so the overlay annotated rows that clients never see.
• It failed generation whenever an annotated identifier was missing from the remote catalog, which could break the automated catalog refresh.

The overlay file, its generator logic, and its tests have been removed. The generator and the tests that existed before this PR are back to main's versions, with only the new validation added.

Validation

• python3 -m unittest discover -s tests: 25 tests, OK in CI.
• python3 scripts/generate_ai_catalog.py passes in CI with 2223 entries; 14 of them carry copilot-connector. ai-catalog.json was regenerated by the workflow.

Follow-up needed in github/agentfinder

Agentfinder's enrichment → search projection currently drops source metadata and replaces source tags with LLM-generated tags, so neither copilot-connector nor metadata.copilotConnector reaches search results yet. Agentfinder needs to:

1. Pass through metadata.copilotConnector (and the source copilot-connector tag) from the ingested entry to search results.
2. Handle data-backed entries such as Kusto (data instead of url). This is an open gap in that path today.