Privacy
This page describes what Lantern Works stores. It matches the code in this repository and the longer notes in PRIVACY.md and docs/MEASUREMENT.md.
The MCP package (lantern-works-deals-mcp)
Telemetry is off by default. With nothing set, the server sends no telemetry and writes no telemetry file. Its only network requests fetch the public feeds, with User-Agent: lantern-works-deals-mcp/<version>.
Set LANTERN_TELEMETRY=1 to opt in to an anonymous tool-call counter. It is enabled only when the value is exactly 1. At most once per UTC day the server sends aggregate counts only: tool name, package version, UTC day and number of calls. It does not send tool arguments, queries or results, or any user, host, client or install identifier. Local counts sit in a cache file until the day is over (at most 7 unsent days). Failures are ignored and never block a tool call.
The collector stores only the total per (day, tool, version). Its rate limiter keeps an in-memory hash and does not write IP addresses to storage. Aggregate totals are public at the collector’s /v1/stats. To turn telemetry off, unset LANTERN_TELEMETRY or set it to anything other than 1.
The hosted MCP endpoint
The hosted worker at https://lantern-works-mcp.lantern-works.workers.dev publishes counts at /stats.json. Call counts are anonymous daily totals. Distinct-caller counts use a daily-salted hash. Weekly repeat-caller counts use a pseudonymous, weekly-rotating id, described below.
- Call counts are totals per UTC day, kind (feed, tool or initialize), path or tool name,
reftag, user-agent class and source class. Those rows do not include an IP address or the raw user agent. - Distinct-caller counts (for
/mcpPOSTs and successful feed fetches) store only a daily-salted SHA-256 hash of the IP address and user agent, together with the same coarse buckets. The raw IP and user agent are never stored, logged or returned. A random salt is created for each UTC day and deleted when that day ends, so hashes from different days cannot be linked. The hashes themselves are kept for 120 days, then deleted. House and CI traffic is not counted. - Weekly repeat-caller counts (for MCP
initializeonly, and only when a Worker secret is set) store a pseudonymous, weekly-rotating id: a 16-character HMAC-SHA256 of the MCP client name, the client version and a coarse network prefix (IPv4 /24 or IPv6 /48), keyed by that secret plus the ISO week, so the same client on the same network keeps the same id within a week, and stored ids cannot be linked across weeks. The raw IP, the prefix, the client name and the client version are never stored, logged or returned. Each stored row holds only the ISO week, the hash and the UTC days that hash was seen; stored ids are deleted after about 18 weeks. Public counts at/stats.jsonare, per ISO week,distinct_callers(how many distinct hashes) andcallers_3plus_days(how many of those hashes were seen on 3 or more distinct UTC days). Our own traffic, CI, known bots, GitHub Actions and MCP Registry health checks (themcp-registryref tag, or a registry-like client name) are not counted. If the secret is not set, or no usable network prefix is available, nothing is recorded and the request is unchanged. - Demand counts (for hosted GPU tool calls) are totals per UTC day, tool, GPU family, region group, price tier and cheapest provider. The GPU family and region group are bucketed to fixed lists: a missing value is stored as
any, and anything else asother. The price tier and provider describe the first result in our own answer, ornonewhen there was none. These rows do not include an IP address, the raw user agent, a caller hash, the tool arguments or any free text. They are kept for 120 days. House and CI traffic and known bots are not counted. - Opt-in price watches: if you create a watch key, the worker stores a hash of the key, the webhook URL, an optional email, your watches and per-key daily counts. Call arguments are not stored.
This website
The site uses Cloudflare Web Analytics for page views; it sets no cookies. Clicks on search, compare, calculator and outbound links are counted as virtual page views under /e/.... Deal links (and the tracked_url field in the feed and MCP results) go through a /go/<deal-id> redirect on the telemetry worker. Vendor links use /go/v/<vendor-slug> (the slug is resolved only from the served GPU feed: a disclosed referral URL, else that vendor’s checked source URL, else the redeem URL of the vendor’s only deal; never a caller-supplied URL). Both add one to an aggregate counter per UTC day, deal id (vendor links are stored as v:<slug>), provider, src and ?ref= tag, then redirect to that listed URL. The counter also keeps a coarse user-agent class (ua_class: browser, bot, script, prefetch or other) worked out from the User-Agent and prefetch headers at request time; the raw user agent is never stored, so automated fetches can be left out of the published click count. No IP address, cookie, user agent or other identifier is stored.