| name | agent-experience-audit |
| description | Audit the MCP server experience from an agent's perspective. Use when checking search quality, response richness, or first-impression for new MCP installs. |
| metadata | {"version":"1.0.0","author":"Master Agent","category":"product-quality","updated":"2026-03-30T00:00:00.000Z"} |
Agent Experience Audit — MCP Dept
You audit what AI agents see when they use Your Project's MCP server. Your goal is to ensure every interaction is obviously better than a raw web search.
Before Starting
- Is production up? Check /health endpoint
- Is SSH available? Try a quick fly ssh command
- What's the context? (New registry listing, post-outreach, routine check)
How This Skill Works
Mode 1: Search Quality Audit
Test the top 5 search queries (auth, analytics, payments, email, database). For each: is the top result relevant? Does it have install_command? Is the tagline useful?
Mode 2: Tool Detail Audit
Pick specific tools (e.g. the 13 we emailed) and check their full detail response. Are descriptions accurate? Are migration signals present? Are compatible tools shown?
Mode 3: First-Install Experience
Simulate a developer who just installed from Smithery/Glama. What's their first query? What do they see? Would they keep using us?
Audit Checklist
Per Search Query
| Check | Pass | Fail |
|---|
| Top result is relevant to query? | Auth → auth tool | Auth → Airflow |
| Top result has install_command? | npm install X | Empty |
| Top result has useful tagline? | Describes what it does | Truncated or generic |
| Category is correct? | Analytics → Analytics | Analytics → AI & Automation |
| Migration signal present? | "14 repos migrated from X" | Empty (acceptable if no data) |
Per Tool Detail
| Check | Pass | Fail |
|---|
| Description > tagline? | Full paragraph | Same as tagline |
| Install command correct? | Matches actual package | Wrong package entirely |
| Migration data present? | Gaining/losing signals | Empty (check sdk_packages mapping) |
| Verified combos present? | "Works with X in N repos" | Empty |
| Trust tier shown? | "New/Tested/Verified" | Missing |
Query via Production
~/.fly/bin/fly ssh console -a your-project -C 'python3 -c "
import urllib.request, json
r = urllib.request.urlopen(\"http://localhost:8080/api/tools/search?q={query}&limit=5\")
data = json.loads(r.read())
for t in data.get(\"tools\", []):
print(t.get(\"slug\") + \": \" + str(t.get(\"migration_signal\", \"no signal\")))
"'
Proactive Triggers
- New tools added to catalog → Check they have install_command before they show in search
- Migration data pushed → Verify signals appear in search results (check Row access pattern)
- Registry listing approved → Run Mode 3 (first-install audit) immediately
- Maker outreach sent → Run Mode 2 on emailed tool pages
Common Issues Found (2026-03-30)
- Plausible Analytics stored as just "Analytics" — no product identity
- ClickHouse miscategorised as "AI & Automation"
- Payment tools (Killbill, Flowglad, dj-stripe) had 0/3 install commands
- Migration signals used row[0] instead of row["column"] — returned NONE for everything
- Cal.com had mailhog install command (wrong tool entirely)