Agent Adoption Intelligence

You measure first-party AI-agent infrastructure for each monitored competitor: MCP server, agent endpoints, Claude Skills, agent toolkit. This skill is a thin wrapper around CompetLab's existing platform scan + a value-add verification pass.

What CompetLab's platform already does (Agent-Adoption Spec — 25 checks)

The platform's start_agent_adoption_scan(domain) runs 25 checks across 4 categories per the open Agent-Adoption Specification:

1. Discoverability — how easy is the vendor to find for AI agents
2. Access control — what AI access permissions are declared
3. Content readability — how AI-friendly is the content surface
4. Agent endpoints — does the vendor declare callable agent surfaces (MCP / API / etc.)

The platform handles the breadth probing. This skill should NOT reinvent that.

What this skill ADDS on top of the platform

1. Protocol-level MCP verification — when the platform detects a candidate MCP URL, this skill verifies via JSON-RPC POST whether it's a real MCP server (vs placeholder / parking page / browser-GET 200). Per PATTERN-url-verification.md § MCP-server-claim verification.
2. Cross-competitor synthesis — the platform returns per-domain results; the skill produces patterns across the monitored set
3. Strategic interpretation — the platform returns raw metrics; the skill produces "this means X for your brand"
4. Optional AI-crawler context — also calls check_ai_crawlers for related-but-distinct signal (how the vendor's robots.txt + meta tags treat AI bots — GPTBot, ClaudeBot, Google-Extended, PerplexityBot)

Workflow

Step 1: Identify competitor set

list_projects → confirm project context
list_competitors → get monitored domain list

Step 2: Fan out platform scans (parallel)

For each monitored competitor domain:

mcp__competlab__start_agent_adoption_scan(domain) → scanId

Optionally also start ai-crawler scan for richer context (related but distinct):

mcp__competlab__check_ai_crawlers(domain, industry: "saas-tech" | "other") → per-bot verdict

Step 3: Poll for completion (5-10s intervals, typical 30-90s per scan)

mcp__competlab__get_agent_adoption_scan(scanId) → status / results

Continue until status === "completed". Collect the 25-check report per competitor.

Step 4: Extract MCP-candidate URLs from each scan result

The platform's "agent endpoints" check category will surface any declared MCP URLs (e.g., from /.well-known/mcp discovery, sitemap detection, or content-readability cues like links to a documented MCP server). Collect these URLs per competitor.

Categorical-zero — CHECK DISCOVERY-VS-CAPABILITY FIRST (banked from real-world validation):

If the platform's well-known-path scan returns ZERO MCP Server Cards across ALL competitors, DO NOT immediately skip Steps 5-6. Per PATTERN-url-verification.md § Categorical-zero case, the absence is on DISCOVERY surface only — capability may still exist via alternative channels. Check FIRST:

1. Content dashboard URLs — does any vendor's sitemap surface URLs matching */mcp-server/*, */mcp/*, */developers/ai/*, */agent-toolkit/*?
2. npm registry — does any vendor publish @{vendor}/mcp-server, {vendor}-mcp, mcp-{vendor} packages?
3. GitHub search — does any vendor have {vendor}-mcp or mcp-{vendor} repos?

If ANY vendor has alt-channel signal → proceed to Steps 5-6 to probe their api.{domain}/v1/mcp, mcp.{domain} etc. via JSON-RPC POST. Real MCP servers often live at API endpoints without /.well-known/mcp mount.

Only after the alt-channel check returns genuinely ZERO signal everywhere → skip Steps 5-6 and record finding as: "Categorical zero MCP adoption in [category]. No JSON-RPC verification needed. First-mover whitespace open."

This is itself the strategic signal — surface in output as "first-mover whitespace" recommendation. Skip to Step 7 (synthesis).

Empirical basis: in a real-world validation run, the well-known-path scan returned 0 MCP Server Cards across 8 vendors (categorical-zero on discovery). One vendor's content dashboard surfaced a documentation page URL pattern. Probing the corresponding api.{vendor}.com/v1/mcp endpoint via JSON-RPC POST returned a JSON-RPC error response (HTTP 401, "Authorization required") — REAL MCP server, auth-gated, not discoverable via well-known path. Without this check, the briefing would have shipped a false categorical-zero claim AND missed the canonical convergence paragraph competitor finding.

Step 5: JSON-RPC POST verification for each MCP candidate URL (THE VALUE-ADD — only run if Step 4 surfaced ≥1 candidate)

This is non-negotiable. Browser GET 200 does NOT prove MCP exists — the URL could be a placeholder, parking page, or real MCP returning HTML to non-protocol clients.

For each candidate MCP URL, use Bash:

curl -sS -X POST "<candidate-url>" \
  -H "Content-Type: application/json" \
  -H "Accept: application/json, text/event-stream" \
  --max-time 15 \
  -d '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2024-11-05","capabilities":{},"clientInfo":{"name":"competlab-verify","version":"1.0"}}}'

Interpret per PATTERN-url-verification.md § Step M2:

Step 6: For verified-real MCP servers, enumerate tool surface

curl -sS -X POST "<verified-mcp-url>" \
  -H "Content-Type: application/json" \
  -d '{"jsonrpc":"2.0","id":2,"method":"tools/list"}'

If auth-gated, the response will say "Authorization required" — that's fine; use mcp__competlab__fetch_url (cleanHtml:true) to read the documentation page (typically at /developers/.../mcp-server/ or mcp.{domain}/docs) and extract the tool list.

Step 7: Synthesize per-vendor + cross-vendor

Per vendor (informed by both the platform's 25-check scan + your JSON-RPC verification):

• Agent-Adoption Specification score (from platform scan)
• MCP server verdict (JSON-RPC verified): REAL / NOT (with reason) / NOT TESTED (no URL surfaced by scan)
• MCP tool surface (if real, list documented tools)
• AI crawler access (from check_ai_crawlers): which AI bots are allowed / blocked
• Other agent surfaces detected by platform (Claude Skills directories, agent-toolkit GitHub repos, etc.)
• Verdict: Production-class / Beta / Reserved-URL-not-shipped / None

Cross-vendor:

• X of N vendors have JSON-RPC-verified MCP servers
• Trajectory comparison to other categories (payments 5/9, CRM 1/4, uptime monitoring 2/9, email warmup 0/12, CS/retention 1/8)
• Strategic gap analysis (who's behind, who's first-mover, where's whitespace)

Output Structure

# Agent Adoption — [Category / Project]
> Generated [date] | Source: CompetLab platform 25-check Agent-Adoption Scan + JSON-RPC POST MCP verification

## Summary
[X of N vendors have JSON-RPC-verified MCP servers; trajectory context]

## Per-competitor
### [Competitor Name]
- **Agent-Adoption Score (platform):** [N/100, breakdown by 4 categories]
- **MCP server:** [REAL — URL + tools + auth] / [NOT — verification reason] / [NOT TESTED — no candidate URL detected]
- **AI crawler access:** [GPTBot allowed/blocked, ClaudeBot allowed/blocked, etc.]
- **Other agent surfaces:** [Claude Skills, agent-toolkit repos, etc. from platform scan]
- **Verdict:** [Production / Beta / Reserved / None]

## Cross-competitor patterns
[Adoption ranking, strategic gap, first-mover identification]

## Verification methodology note
[Platform's 25-check scan was Stage 1; JSON-RPC POST verification was Stage 2 value-add]

What NOT To Do

• Don't bypass the platform scan and DIY-probe candidate URLs. The platform's 25-check scan is comprehensive; this skill ADDS verification on top, doesn't replace.
• Don't report platform-scan "MCP detected" as verified MCP. Detection is Stage 1; JSON-RPC POST verification is Stage 2. Both required for a real claim.
• Don't include AI-crawler signals as "agent adoption." They're related but distinct. AI crawlers consume content; agents call actions. Surface both, label clearly.
• Don't list `/llms.txt` as agent infrastructure. llms.txt is an AI-CRAWLER preference declaration (similar to robots.txt for LLMs), NOT an agent endpoint. Different concept. If /llms.txt is interesting for the briefing, it belongs in the AI-crawler readiness signal via check_ai_crawlers, not agent-adoption.

Decision Questions

1. "Should we surface our MCP server presence in homepage messaging, or keep it as developer-only surface?" — affects positioning recommendations
2. "What's our agent-developer hiring posture — do we have engineers who can ship an MCP server in 1-2 weeks?" — affects timeline recommendations
3. "Are we positioning as an AI-agent-native platform OR as a layer/tool that complements existing platforms?" — affects which MCP tool surface to expose

Error Handling

• Underlying scan returns no result or times out for a competitor: record as "agent-readiness signal unavailable for {vendor} in this run." If the broader pattern suggests bot-protection (other dimensions also returning thin data for the same vendor), surface that as a customer-facing finding — bot-protection in 2026 means invisible to AI training crawlers + agent toolkits + comparison-engine bots, which is itself competitive intelligence about the vendor's discoverability posture.
• JSON-RPC POST returns 5xx: retry once; if persistent, flag as inconclusive
• GitHub API rate limit (60/hr unauth): mention in output; suggest authenticated mode if user has GH_TOKEN

Empirical reference

Internal validation runs demonstrated this skill's methodology against real vendor URLs:

• REAL MCP server confirmed: documentation page detected at {domain}/developers/ai/mcp-server/; JSON-RPC POST initialize against the actual API endpoint api.{domain}/v1/mcp returned {"jsonrpc":"2.0","error":{"code":-32000,"message":"Authorization required"}} (HTTP 401) — JSON-RPC error body proves the server speaks the protocol.
• FALSE-POSITIVE 1 caught: vendor reserved mcp.{domain} subdomain (GET returned 200 + "Domain not found" HTML); JSON-RPC POST returned 404 + "Cannot POST /" (Express error) → NOT MCP.
• FALSE-POSITIVE 2 caught: vendor reserved developers.{domain}/mcp URL (GET returned 200 + empty HTML); JSON-RPC POST returned 403 + CloudFront "method not allowed" → NOT MCP (CDN-fronted cache-only path).

Without the JSON-RPC verification step, 2 of 3 MCP claims would have shipped to customer briefings as false positives. The skill's value-add is real.