Agent Skills: ecosystem-usage

Use this skill when asked about MCP adoption, AI-agent traffic, raw Markdown consumption, curl usage, or which clients consume the evlog docs. Web Analytics sees browsers only; this skill measures the requests that never run a pageview script.

Source of truth

Use the read-only vercel__create_observability_query tool directly.

• Metric: vercel.request.count, aggregation sum.
• Scope: type: 'project', ownerId: the evlog team id, projectIds: the docs site project id (both pre-scoped in the connection description).
• Always filter to environment eq 'production'.
• Use ISO UTC timestamps for startTime and endTime.
• Because the result is read against a comparison, always query the requested window and the immediately preceding equal-length window with the same scope and filter, ungrouped.
• A tool-call concurrency limit is not a total-query budget: send further read-only calls until every requested metric is collected.

Query recipes

Run independent queries in parallel, ungrouped first for the exact total, then grouped for the breakdown.

1. Agent-facing total:
   • Filter: environment eq 'production'
   • Group by client_user_agent (limit 25) and bot_category + bot_name (limit 20).
2. Explicit Markdown URLs:
   • Filter: endswith(request_path, '.md') and environment eq 'production'
   • Group by request_path (limit 10) and client_user_agent (limit 10).
3. Content-negotiated Markdown:
   • Filter: contains(http_accept, 'text/markdown') and environment eq 'production'
   • Group by request_path (limit 10) and client_user_agent (limit 10).
4. Agent discovery and intake:
   • Filter: (request_path eq '/llms.txt' or request_path eq '/llms-full.txt' or request_path eq '/sitemap.md' or request_path eq '/.well-known/mcp/server-card.json') and environment eq 'production'
   • Group by request_path (limit 10), client_user_agent (limit 10), or bot_category + bot_name (limit 10).
   • Keep these separate from content reads: fetching an index does not prove the client consumed a documentation page.
5. curl traffic, only when explicitly asked:
   • Filter: contains(client_user_agent, 'curl/') and environment eq 'production'
   • Group by request_path and client_user_agent; exclude asset paths from the interpretation.

Interpretation rules

• Call the result HTTP requests, never tool calls, sessions, users, or unique agents. Initialization, discovery, tool calls, retries, and notifications each count separately.
• Use the ungrouped summary as the authoritative total. Do not add grouped rows or timeseries buckets to reconstruct it.
• Empty or generic user agents (node, undici, Go-http-client, python-httpx) identify a client stack, not an agent product. Never rename a generic or empty user agent into a specific product.
• Report at most five recognized product rows with exact counts, then at most three generic stack rows, then the empty-user-agent row when present. Never sum version variants.
• A .md path or a curl user agent alone does not prove AI usage: humans use "View as Markdown", scripts use curl. Treat explicit Accept: text/markdown, known AI bot categories, and MCP transport paths as the stronger signals.
• Top-N grouped rows are partial: describe them as top returned rows, never as all traffic.
• If a response says truncated: true or reports truncation.omittedArrayItems, only the returned timeseries was shortened; report the summary total and do not call it a data gap. Only label a real data gap when the API explicitly reports one after truncation is ruled out.
• If a query times out, shorten the window or drop a high-cardinality grouping; the ungrouped total stays authoritative.
• Browser traffic stays with vercel__count_pageviews and vercel__aggregate_pageviews; label it as browser pageviews and never present it as total readership when agent-facing traffic is in scope.

Output

• Include the exact requested time window and every requested metric with its HTTP request count.
• When a number is shown against a comparison, take the change from the queried preceding window; if that query was not performed or failed, write "change unavailable: <reason>" instead of implying a delta.
• If a required query failed, show that metric as unavailable beside the successful totals, with the concrete error in one line.
• End with one short caveat that HTTP request volume is not logical tool-call volume.