JSON Schema App MCP: Claude Code Plugin

Connects Claude to the JSON Schema App MCP server (https://mcp.jsonschemaapp.com/mcp) so you can check structured data (JSON-LD / Schema.org), scan your site, audit AI-crawler access, and read your structured-data report, right from Claude.

This plugin is a client: it only connects to the hosted MCP server. It does not change or redeploy any server code, so nothing in the existing MCP/app flow is affected.

Quick start

/plugin marketplace add MakkPresstech/json-claude-plugin
/plugin install jsonschemaapp-mcp@jsonschemaapp
/jsonschemaapp-mcp:report

The first command that calls a tool opens the connect page in your browser. See First connection.

Contents

• What's inside
• MCP tools exposed by the server
• First connection
• Authentication
• Install
• Fewer permission prompts
• Usage
• Example output
• Skills (model-invoked)
• Consistent, grounded audits
• Troubleshooting
• Note on FAQPage / HowTo

What's inside

json-claude-plugin/
├── .claude-plugin/
│   ├── plugin.json              # Plugin manifest
│   └── marketplace.json         # Marketplace entry so it installs via /plugin
├── .mcp.json                    # Connects to https://mcp.jsonschemaapp.com/mcp (OAuth)
├── commands/                    # Slash commands (one per workflow)
│   ├── check-page.md            # /jsonschemaapp-mcp:check-page   -> check_page_schema
│   ├── scan-site.md             # /jsonschemaapp-mcp:scan-site    -> scan_site + status + result
│   ├── report.md                # /jsonschemaapp-mcp:report       -> get_report
│   ├── check-llms.md            # /jsonschemaapp-mcp:check-llms   -> check_llms_txt
│   └── ai-search-audit.md       # /jsonschemaapp-mcp:ai-search-audit (full audit, all tools)
├── agents/
│   └── schema-auditor.md        # Subagent that drives a full audit end to end
├── skills/                      # Model-invoked skills (auto-activate from natural language)
│   ├── structured-data-audit/   # Full audit workflow over the MCP tools
│   ├── schema-rich-results/     # Which Schema.org types earn rich results (FAQPage/HowTo rule)
│   ├── ai-search-optimization/  # llms.txt + AI-crawler access via check_llms_txt
│   ├── draft-llms-txt/          # Draft a valid llms.txt from scan data (client-side)
│   ├── ai-crawler-access/       # robots.txt snippets to allow AI crawlers (any platform)
│   └── audit-report-format/     # Canonical report shape (verdict + P1/P2/P3 + tool citations)
├── hooks/
│   ├── hooks.json               # SessionStart connectivity reminder
│   └── connectivity-check.sh    # Non-blocking "connect the server first" notice
├── docs/
│   └── images/                  # Banner, demo GIF and screenshots used in this README
├── icon.png                     # Plugin icon for the directory listing
├── LICENSE                      # MIT
└── README.md

How the pieces fit together

flowchart LR
    subgraph Plugin["json-claude-plugin (client)"]
        CMD[Slash commands]
        SK[Skills]
        AG[schema-auditor subagent]
    end
    CMD --> MCP
    SK --> MCP
    AG --> MCP
    MCP["mcp.jsonschemaapp.com/mcp"] --> T1[check_page_schema]
    MCP --> T2[scan_site / get_scan_status / get_scan_result]
    MCP --> T3[check_llms_txt]
    MCP --> T4[get_report]

MCP tools exposed by the server

Webflow tenants also get server-side starter prompts (optimize_for_ai_search, fix_ai_crawler_access, check_schema_coverage, assign_schema_template) which surface automatically in Claude's prompt picker when connected as a Webflow store.

First connection

You need a JSONSchemaApp account and a pairing code. No account yet? Create one free, install the app on your store, then continue below.

1. In Claude Code, run any command, for example /jsonschemaapp-mcp:report. Your browser opens the Connect an AI assistant page.
2. In another tab, open JSONSchemaApp from your store admin (or sign in at app.jsonschemaapp.com), go to Settings, then AI assistant access, and click Generate pairing code.
3. Copy the 8-character code, paste it on the connect page and click Allow. You are sent back to Claude and the command continues.

The connection is read-only: Claude can read your schema reports and scan results but cannot change your store. Disconnect anytime from Settings, AI assistant access.

sequenceDiagram
    participant U as You
    participant C as Claude Code
    participant M as mcp.jsonschemaapp.com
    participant A as JSONSchemaApp dashboard
    U->>C: /jsonschemaapp-mcp:report
    C->>M: Tool call (no token yet)
    M-->>C: 401 + WWW-Authenticate
    C->>U: Opens browser to the connect page
    U->>A: Settings, AI assistant access, Generate pairing code
    U->>M: Paste code, click Allow
    M-->>C: OAuth token
    C->>M: get_report
    M-->>C: Report data

Authentication

The plugin's only built-in auth path is OAuth. There is nothing to configure in the plugin itself; the only manual step is the pairing code above.

.mcp.json points at the HTTP endpoint with no credentials, so on first use Claude discovers the OAuth flow automatically (via the server's 401 + WWW-Authenticate and /.well-known/oauth-protected-resource) and prompts you to connect.

Using an API key instead

If you prefer a jsa_live_... key, add your own user-scope MCP server alongside the plugin. The plugin does not ship a key option, because Claude Code disables OAuth fallback for any server whose config sets an Authorization header, which would take the OAuth path away from everyone who does not use a key.

claude mcp add --scope user --transport http -H "Authorization: Bearer YOUR_API_KEY" jsonschemaapp-key https://mcp.jsonschemaapp.com/mcp

Replace YOUR_API_KEY with your own key. Two things to note:

• Pick a server name other than jsonschemaapp. The example uses jsonschemaapp-key. Reusing the plugin's own server name shadows it, and you then get confusing connection errors because both entries claim the same name.
• Your key lives in your personal Claude Code config, not in this repository. Keep it out of version control, and revoke it in your dashboard under Settings if it leaks.

With a key configured this way, disable the plugin's own server in /mcp so you are not connected twice.

Install

Via marketplace (recommended). This directory ships a .claude-plugin/marketplace.json, so it installs with /plugin:

/plugin marketplace add MakkPresstech/json-claude-plugin
/plugin install jsonschemaapp-mcp@jsonschemaapp

Local dev. Load the directory directly without a marketplace:

git clone https://github.com/MakkPresstech/json-claude-plugin.git
claude --plugin-dir ./json-claude-plugin

Verify the connection with /mcp (the server lists as plugin:jsonschemaapp-mcp:jsonschemaapp when installed as a plugin), then run /jsonschemaapp-mcp:report to confirm the tools respond. On session start the plugin's connectivity hook prints a short reminder to complete OAuth first. See Fewer permission prompts.

Fewer permission prompts

Each mcp__plugin_jsonschemaapp-mcp_jsonschemaapp__* tool call prompts for approval by default. To let the six read-only tools run without interruption, add them to your project .claude/settings.json (or user settings). Nothing here changes the server. It only pre-approves client-side tool calls you'd otherwise approve by hand:

{
  "permissions": {
    "allow": [
      "mcp__plugin_jsonschemaapp-mcp_jsonschemaapp__check_page_schema",
      "mcp__plugin_jsonschemaapp-mcp_jsonschemaapp__scan_site",
      "mcp__plugin_jsonschemaapp-mcp_jsonschemaapp__get_scan_status",
      "mcp__plugin_jsonschemaapp-mcp_jsonschemaapp__get_scan_result",
      "mcp__plugin_jsonschemaapp-mcp_jsonschemaapp__check_llms_txt",
      "mcp__plugin_jsonschemaapp-mcp_jsonschemaapp__get_report"
    ]
  }
}

Usage

/jsonschemaapp-mcp:report                       # current state of your store
/jsonschemaapp-mcp:scan-site 200                # scan up to 200 pages, then summarize
/jsonschemaapp-mcp:check-page https://site/x     # check one page's structured data
/jsonschemaapp-mcp:check-llms example.com        # llms.txt + AI-crawler access
/jsonschemaapp-mcp:ai-search-audit               # full readiness audit

Or just ask in natural language (e.g. "audit my store's structured data") and the schema-auditor subagent will drive the tools for you.

Example output

A shortened /jsonschemaapp-mcp:ai-search-audit result for a sample store:

Verdict: Product schema is solid, but AI crawlers are blocked and llms.txt is missing.

P1 (blocking)
- robots.txt disallows GPTBot and ClaudeBot on all paths.          [check_llms_txt]
- 14 product pages missing "offers.price" in Product schema.        [get_scan_result]

P2 (high value)
- No llms.txt at example.com/llms.txt.                              [check_llms_txt]
- Organization schema has no "sameAs" social profiles.              [check_page_schema]

P3 (incremental)
- 6 blog posts use Article without "dateModified".                  [get_scan_result]

Scanned 120 pages, 98 with valid JSON-LD. Plan: Pro, 1,840 AI credits left. [get_report]

Skills (model-invoked)

Unlike the slash commands (which you trigger explicitly), the bundled skills load automatically when your request matches their description. No command needed:

They carry the same workflow and rules as the commands/subagent (including the FAQPage/HowTo rich-result rule) so behavior stays consistent however you invoke it.

Consistent, grounded audits

So every audit reads the same way and stays efficient, the commands, subagent, and skills share three conventions (all client-side prompt guidance, no server change):

• Standard report format. Every answer leads with a one-line verdict, then findings grouped P1 (blocking) → P2 (high value) → P3 (incremental), each citing the tool that produced it. Defined once in the audit-report-format skill. Findings must trace to real tool output, with no invented scores or errors.
• Poll backoff for scans. After scan_site, get_scan_status is polled with a backoff (first poll ~5s, then 5s → 10s → 20s → 30s, ~30s thereafter) instead of a tight loop.
• Pagination discipline. get_scan_result is read with only_issues: true and limit: 50, paging via offset, so large scans don't flood context; the full inventory (only_issues: false) is pulled only when actually needed.

Troubleshooting

Note on FAQPage / HowTo

FAQPage and HowTo markup no longer produce Google rich results. The tools report them at the ai_readability level (useful for AI assistants). The commands and the subagent are instructed never to recommend them as Google rich-result wins.