Divar MCP is a read-only public service. There is nothing to log in to and no user data is stored.

• All 7 tools are read-only. No tool can change, delete or publish anything.
• No API keys are needed to use the hosted endpoint.
• Nothing is persisted server-side. The only state is a short-lived response cache (per-isolate, plus Cloudflare's edge cache) that other isolates cannot read.
• Phone numbers need the seller's own login - this server never logs in and never returns them.
• Data comes from Divar's public web API, which is undocumented and can change without notice. This project is not affiliated with or endorsed by Divar.

The public endpoint is guarded by a per-IP budget on POST /mcp, counted in two places:

1. At the edge, on the hosted deployment - Cloudflare's rate-limit binding ([[ratelimits]] in wrangler.toml) counts every POST per client IP, and its counter is shared across the isolates of an edge location - which is why a burst cannot get around it by landing on many isolates. It answers 429 with a retry-after header.
2. In the server code, wherever it runs - the same budget as an in-code sliding window (per isolate, held in memory). It is the whole guard for a deployment made from this source (a plain wrangler deploy carries no binding unless you add one) and for the Node --http transport, and it is what the Worker falls back to if the edge counter itself fails - a guard that breaks the service it guards is worse than no guard.

Notes:

• /health, / and the preflight (OPTIONS) are not counted by either counter above, so dashboards and browser clients are not punished by this service.
• 20 a minute is ~3x what a normal agent session uses (requests to Divar are paced near one tool call a second), so ordinary use never notices it; bursts of parallel tool calls can.
• If you are rate-limited, back off until retry-after says you may return - hammering while refused only keeps the count up.
• A third 429 exists that neither counter produces: Cloudflare's platform can refuse the whole worker briefly with a plain-text body (error code: 1027), no retry-after, on every path including /health. Observed once, 2026-10-04, ~30 minutes, then it cleared by itself; it does not come from this code (the GET paths never touch the counters above). If you hit it: there is no header to read - wait a few minutes and retry once. The lever lives in Cloudflare's dashboard (account → Security / Logs for this worker), not in this repository.
• The edge counter is per edge location rather than one global number, and that is enough: a client whose calls land on different isolates still lands in the same place. Measured on the live endpoint 2026-09-30: 30 POSTs over one connection - what a real MCP client does - passed 20 and then answered 429 for the rest.
• A deployment behind its own domain can add a third layer - a Rate Limiting rule (dashboard: Security → WAF → Rate limiting rules). The hosted *.workers.dev endpoint cannot carry one, because its zone belongs to Cloudflare.
• The in-code window holds its counters in memory rather than in KV: sharing them through KV would burn the free write quota in minutes, and for abuse-throttling a per-isolate window is enough once the edge counter above is doing the global work.

The endpoint is keyless and read-only, so it sends Access-Control-Allow-Origin: * and answers OPTIONS preflights. There are no cookies or credentials to leak; browser-based MCP clients (Inspector, web agents) need this to connect at all.

Please do NOT open a public issue. Report privately via the Security tab (Advisories → Report a vulnerability).