MCP Server

NSIN hosts a remote Model Context Protocol (MCP) server — ask your AI assistant to "purge /assets on example.com" or similar, in plain language.

• Endpoint: https://api.nsin.ir/mcp (Streamable HTTP)
• Auth: your API key as Authorization: Bearer nsin_…
• Nothing to install — NSIN hosts it.

Connect

Claude Code

claude mcp add --transport http nsin https://api.nsin.ir/mcp \
  --header "Authorization: Bearer nsin_your_key_here"

Claude Desktop

Add this to claude_desktop_config.json (Settings → Developer → Edit Config). It uses mcp-remote to send your key as a header:

{
  "mcpServers": {
    "nsin": {
      "command": "npx",
      "args": ["-y", "mcp-remote", "https://api.nsin.ir/mcp",
               "--header", "Authorization: Bearer nsin_your_key_here"]
    }
  }
}

Cursor

Add to .cursor/mcp.json (or global ~/.cursor/mcp.json):

{
  "mcpServers": {
    "nsin": {
      "url": "https://api.nsin.ir/mcp",
      "headers": { "Authorization": "Bearer nsin_your_key_here" }
    }
  }
}

Tools

✓ means a read-only key can use the tool. Read-only keys get a smaller list — write tools are never shown to the assistant. Every tool also tells the client whether it only reads, and whether it deletes or replaces something, so Claude and other clients can ask you before a risky call.

Domains and DNS records

Zone import and export

An NSIN export imports back as it was: proxied records stay proxied, and the apex alias comes back as ANAME.

SSL

Sharing and gateways

Rules

kind: cache, waf, drop, rate_limit, captcha, bot_route, fingerprint, redirect, rewrite, header, origin_pool, origin_route, basic_auth, error_page or optimize. Rules only affect proxied hostnames. A kind your plan does not include saves switched off and cannot be switched on.

Cache

Traffic and logs

Uptime

Email Routing

A destination forwards nothing until its owner clicks the verification link.

Account

Writing rules

Call get_rule_schema with the kind first: it lists every field of that kind's body with its type, allowed values, defaults and limits, plus example bodies. Then create_rule with kind and the body as rule. A field the kind doesn't have is refused before anything is saved.

update_rule takes a partial body: fields you leave out keep their values, and a list you send replaces the old one. To stop a rule without deleting it, send {"enabled": false}. Lower priority runs first; reorder_rules sets the order of one kind in one call.

Purging the cache

Three purge tools, most to least precise. All need cache.edit permission and a plan with cache purge.

purge_cache_by_path — the everyday, after-a-deploy tool. path: "/assets/*" (a directory), path: "/" (homepage), path: "/blog/post-1" (one page). * is a wildcard, no regex.

• No * = substring match — /logo.png also matches /img/logo.png?v=2 — anchor it (/assets/*) when that matters.
• Optional hostname/node narrows it; omit both = every hostname, every node.
• All cache-key variants (query strings, mobile, WebP/AVIF, CORS) purge too.
• Up to 10,000 objects per call; truncated: true means repeat. Preview with list_cache_entries.

purge_cache_entries purges exactly the rows list_cache_entries returned — for over-matching patterns.

purge_cache drops everything cached, on every node — the blunt option, origin takes all traffic until refill.

Targeted tools also accept mode: "refresh": drops the cached copy, keeps the entry listed; the next visitor re-fetches from origin. Deploying? Use set_developer_mode, not a purge loop.

Custom analytics queries

run_analytics_query covers what no dedicated tool does — "top 20 URIs by 5xx count, last 7 days". Pass a ClickHouse SELECT over the requests table, get columns and rows back as JSON, capped at 10,000 rows.

Sandboxed:

• Read-only. Only a single SELECT/WITH … SELECT. Writes, DDL, settings changes, INTO OUTFILE/FORMAT, table functions (url, file, remote, s3, …) are rejected.
• Your data only — scoped to domains your key can access, even in a subquery.
• Bounded execution time and rows.

Useful requests columns: event_time, domain_id, hostname, method, uri, status, bytesIn, bytesOut, duration (ms), remoteAddr, country, asn, asnOrg, userAgent, cacheStatus (hit/miss/bypass), bypassReason, cacheAge, reqStatus (cache/proxied/direct), isWS, protocol, referer, rayId, node, originStatus, originAddr, error, errFault (origin/edge/client), errReason, upstreamDuration (ms), originConnReused, originDialMs, originTlsMs, originTtfbMs, originTlsVersion, ja4, botKind, wafScore, mdConverted, optKind. Bucket with toStartOfHour(event_time) or toStartOfDay(event_time).

errFault = 'origin' catches outages with no 5xx status — timeouts, refused connections, hangups — that status >= 500 misses.

Permissions and safety

• MCP uses your account's own permissions — Viewer on a shared domain blocks write tools there, even on a full-access key.
• Plan features gate too: monitoring, logs, WAF, cache purge; a tool needing one fails with the reason.
• Every action gets the REST API's validation, audit logging, usage metering, and counts toward the 300 req/min rate limit.
• Tools that delete something for good take confirm: true; the assistant should ask you first.
• No key can move money: top-ups, plan purchases and choosing a domain's billing member stay in the panel.
• Revoking the key in the panel disconnects the assistant.