Privacy
What this service records when you use the website, the REST API or the MCP server, and what it does not.
Updated 2026-11-14
No account, and one small cookie on the analysis page
There is no account, sign-up or API key. The only cookie this service sets is a functional one on an analysis page (/c/<id>): when your browser opens that page as a person, the page sets a cookie named pf_ok_<id> with the value 1, so that reloading that page shows the full page again. It lasts 1 day, is sent only to that page (path /c/<id>), is HttpOnly and SameSite=Lax, and holds nothing about you. It is not used for tracking, analytics or advertising, and nothing else on this service sets a cookie. There are no analytics or advertising tags. The only scripts are this site's own: the canvas page's live update, and on the canvas, search and answer pages the WebMCP tools an AI in your browser can use.
What is recorded
- One log line per request: the time, the HTTP method, the path without the query string (a product page or a REST plan address names that product or plan, and a purchase link, which goes through /go/ to the provider, names the plan whose link was followed, with the response token the link carries: 8 random characters made for one response that hold nothing about you, and the place of that plan among the plans the response listed, and its fit score in that response (a number from 0 to 100 counted from the conditions the plan met, shown as fit in the plan), and the mark via=app when the link was followed from the comparison card that some assistants show for a search result; no other value of via is recorded), the User-Agent header, the Accept header (the content types the client asked for, at most 200 characters), the media type of the response (for example text/html), whether the User-Agent names Claude Code or is a Claude-User caller of another kind, the response status and how long the request took. A request refused by the rate limit is marked as such, a request to a verdict address refused as automated (a crawler that robots.txt closes it to, or a program behind a browser User-Agent that sends no Sec-Fetch-Mode header) is marked with that reason, a request to a verdict address or a purchase link keeps the values of the three Sec-Fetch headers the browser sent (mode, destination and whether you started it, at most 16 characters each), and a server error keeps the first lines of the error.
- The server also keeps in memory, for up to seven days and never in a response, what each response token was made for: the time, whether it was a page, the REST API or the MCP server, the tool, the client name the caller gave, the plans in the order the response listed them and the fit score of each. It is used only to give a followed link its place and its score in that list. A restart clears it. The log line of a request whose response carried plan links also holds that response token, the number of plans it linked, whether it was a page, the REST API or the MCP server, and the tool and the fit score and metrics of each plan the response listed (numbers worked out from the conditions and the request, not from you), the answer shape the response was served in (a letter; empty on this public server) and how many official source sentences it carried, so that the responses a followed link came from, and the scores of the plans that were followed, can be counted. The token is random and holds nothing about you; the plans themselves and the workload are not in it.
- For web page requests also: the Referer header (the page that linked here), reduced to its origin and path, without the query string; it is kept for the same time as the rest of the log.
- For every request also: whether an Accept-Language header came (1 or 0; the languages in it are not kept), because a person's browser sends one and many scripts do not.
- For every request also: a visitor code, 16 characters computed from your address and User-Agent with a secret key. The key is made in the server's memory when the day changes (Korea time) and when the server starts, is never written down, and is thrown away when the day ends. The code cannot be turned back into your address, and the same browser gets a different code the next day (and a second one on the day the server restarts). It is used to count visitors per day and to tell whether a purchase link was followed by someone who had read the page the link was on. It is kept for the same time as the rest of the log. Requests that carry the same code with no more than 30 minutes between them are read together as one visit, to see how a visit goes (which pages, whether a purchase link was followed, which tools were called). There are still no cookies and no scripts; the visit is worked out from the log afterwards.
- For every request that carries one: the Signature-Agent header, as sent (at most 200 characters; an AI agent that signs its requests names itself in it, for example https://chatgpt.com), so requests from an AI agent are not counted as a person's browser.
- For every request that carries one: the Sec-Purpose (or Purpose) header, as sent (at most 40 characters; a browser that loads a link ahead of a click says prefetch in it), so a link loaded ahead is not counted as a click.
- For MCP requests also: the MCP protocol headers (version, method and name), the Accept header, the client name and version the client states in its initialize request (for example claude-code and 2.1.281; at most 64 characters each), whether a session header was sent (as two short codes: one that changes every day, and a hash of the session id, 16 characters, so that the calls of one MCP session can be read together; neither can be turned back into the header), whether the request was refused for its Origin header, the JSON-RPC method, the tool name, and whether the tool returned an error, with its error code and the names (not the values) of the fields that caused it.
- For REST API requests also: the X-Plan-Finder-Client header if it is sent (a client name such as skill/0.1.1, at most 40 characters).
- For a tool call (MCP) or a REST API call also: which kinds of API the workload asked about (for example web scraping, search, browsers, email, speech, document parsing) and the names of the conditions given as required or as preferred (for example needs_zdr or min_concurrency). Only names from a fixed list are kept; the values (numbers, countries, budgets, plan names) and any other key are not. For a plan search, a plan check or an analysis report that carries the user's words (the question), the words themselves are kept, cut to 300 characters, with email addresses, phone numbers, long digit runs and the query part of a web address replaced (for example [email]), and with the field names their cue words pointed to. For an analysis report the names of the fields read from the words are kept too, and the words of the question that no reading rule matched (single words, masked the same way, at most 300 characters). Whether the answer carried the address of an analysis page is kept as yes (analysis_link), and when an analysis report asked the person for more, only the names of the missing items (for example product_type or monthly_audio_hours), never what the person answered.
- For an action on an analysis page (canvas_do): the action's own settings (conditions changed or added, plan ids excluded, highlighted or compared, the sort, a note cut to 300 characters with email addresses and phone numbers replaced), and for a search made again on the same page, the names of the conditions added, changed or removed.
- The sentence an assistant passes on as the person's request (the optional question), and what an assistant passes on with send_feedback, are kept for the same period as this log, to learn what people look for and which information is missing and add it. Statistics grouped from them (requests per kind of API, the share of each condition, the plans chosen, the kinds of reason for leaving, what assistants did on an analysis page: conditions added, plans excluded or highlighted) may be shared with or sold to providers and others as a report. The text of a request, the visitor code and single requests are not put in a report and are not sold, and a group of fewer than 10 requests is left out of a report.
- For a plan search or a plan check also: which plans were judged, which plan came first among those that meet every condition, and for each plan the names of the required conditions it did not meet or could not confirm, with the volume rounded down to 1, 2 or 5 times a power of ten (for example 37,000 is logged as 20,000). Plan ids and condition names only; no value you gave, no text and no personal data.
- For a tool call or a product page that carries optional parts (the cheapest plan of another provider, what differs between the plans checked, the fields a question may have left out, the one result link, the verdict section of a product page, the verdict links of a vendor page, the price and limit changes of a product page) also: the names of the parts the response carried (shown) and, when a later call on the same path names a plan, condition or field only such a part had shown, or a request arrives through the link such a part gave (a page or verdict address with the channel tag share or page), the name of that part (took). The path is the MCP session code or the visitor code; what it was shown is kept in memory for 30 minutes and dropped when the UTC day changes. Part names only; no plan id, condition, field or value is logged.
- For a request_coverage call (a model telling us a field, provider, condition or plan it missed) also: the kind, the name (at most 80 characters) and the optional reason (at most 200) the caller wrote. The call takes no name, e-mail address or link and refuses them; it has no reply and is read in aggregate on the board.
- For a send_feedback call (an assistant passing on what its user said of a result, or a provider's assistant about its own plans) also: who sends (user_assistant or provider), the kind, and the optional provider name (at most 80 characters), official page address (https, at most 300), offer id, subject and message (at most 300) the caller wrote. The call takes no person's name, e-mail address or other link and refuses them; it has no reply, changes no published fact, and is read on the board.
- For a page request whose address carries a channel tag from a short fixed list (registry, smithery, glama, mcpso, readme, guide, answers, try, webmcp, llms, share, limits, plan, planform, page, vendor, vs, canvas, results, product, watch, rank, history, magazine, maghome), that tag, so the places people arrive from can be counted.
- For the analysis canvas (a page an AI assistant can open for you with its comparison of plans): the log line of a canvas request holds the canvas id (12 random characters that hold nothing about you), the plan id when a plan button on it was pressed, and, when the assistant changed it, the name of the change from a fixed list (for example sort or compare); the line of a plan search or check answer holds whether the answer carried a canvas link. The canvas itself, unlike the log, keeps what the assistant sent to make and change it (the conditions, the goal, must-haves and context it gave, and the notes and explanations it wrote) and the plans it shows, in a file on the server, for at most 30 days after it was opened. Anyone with its address can open it, so an assistant should not put personal details in it.
- For an install button (Add to Claude, ChatGPT, Cursor, VS Code or Claude Code, under a verdict): which client it was for, so presses can be counted per client and page.
- When a link you followed names an AI assistant as its source (utm_source, such as chatgpt.com): that assistant's name from a fixed list (ChatGPT, Perplexity, Claude, Gemini, Copilot), and whether a verdict link opened in your browser was sent on to the search page.
- Not recorded in the log: the values of your workload, conditions and other inputs (tool arguments, request bodies and query strings), and your IP address; the log keeps only the network that address belongs to (its AS number and name) and the country, never the address itself.
- To apply rate limits, the server keeps the address a request came from in memory. It is not written to the log or anywhere else. It stays in memory until the server restarts, or until 10,000 addresses are held, when the idle ones are dropped.
How it is used and shared
The log is used to keep the service running and to count how often pages and tools are used. It is not sold, not shared with third parties and not used for advertising.
Purchase links pass through this site's /go/ address, which records the request like any other and forwards it at once, and lead to the provider's own site, where the provider's privacy policy applies. The address it forwards to carries utm_source=api-plan-finder and utm_medium=referral, which name this service and hold nothing about you. A /d/ address, which a search or check result gives as next_action, returns the result of the request its response token was made for; the server keeps that request (the conditions asked, no words of a question) with the token in memory only, for seven days. Whether a link is an affiliate link is stated on the disclosure page. When an affiliate program takes a sub-ID, the response token goes with the link to that program, so a sale can be matched to the response it came from; the token holds nothing about you.
How long it is kept
The request log is kept in one file per day (UTC), and each file is deleted 30 days after its day began. The check runs when the server starts and every hour.
Contact
https://github.com/scraping-api-plans/web-data-plan-finder/issues