Last updated: November 19, 2026
HefestoAI is operated by Narapa LLC, based in Miami, Florida. Contact: support@narapallc.com
This covers the MCP endpoints https://hefestoai.narapallc.com/api/mcp-protocol and https://hefestoai.narapallc.com/api/mcp-core and the analyzer behind them, https://hefestoai.narapallc.com/api/analyze. Neither one needs an account, an API key or a cookie.
What we receive. The JSON-RPC request. For the analyze tool this includes the file paths and file contents (or the code snippet) you send, plus the severity and format options. Like every web request, it also carries your IP address and request headers such as User-Agent.
What happens to your code. The analyzer writes the files to a new temporary directory, runs the open-source hefesto-ai engine on them, returns the findings and deletes the directory when the request ends. Submitted code is not written to a database, not written to our logs, not used to train models and not shared with anyone except our hosting provider, which runs the analysis.
Rate limit and your IP address. The analyze tool is limited to 10 requests per minute per IP address. The MCP endpoint passes your IP address to the analyzer for this check. The shared counter is kept in Upstash Redis under a key made from a SHA-256 hash of a fixed prefix plus your IP address and the current minute, and it expires after 120 seconds. Anyone who knows an IP address can compute the same hash, so we treat the hash as personal data. If Redis cannot be reached, the server counts in its own memory, keyed by the IP address itself. That counter is never written to disk and is cleared when the server instance restarts or when it holds 10,000 addresses.
Logs. For each analyze request our code logs which rate-limit backend answered. That line contains no IP address, no code and no findings. Our MCP code no longer logs the Origin host, the User-Agent or the MCP-Protocol-Version header of each request: that temporary line has been removed. Vercel also records basic request details (time, path, method, status, user agent, region). On our Vercel plan, runtime logs are kept for 1 hour.
Usage counts. When enabled, the MCP server sends aggregate events to PostHog (US cloud by default). Each event has the event name (initialize, server/discover, ping, tools/list, tools/call, tools/call_unknown_tool, isError or rate_limited), the tool name, the outcome, the protocol version, and the client name and version the client reports (in the initialize request, or in a request's own client information). A failed tool call also carries an error type from a fixed list (for example timeout or invalid_path). A call to a tool that does not exist is counted as tools/call_unknown_tool, without the name that was requested. That event also carries requested_tool_class, a label from a fixed list (case_or_prefix_match, empty or other): case_or_prefix_match when the name matches a real tool except for capitalization or an added prefix or suffix, empty when the name is missing or blank, and other otherwise. The requested name is not part of the event. When analyze is rejected before it runs, or the analyzer answers HTTP 400, the event also carries error_reason, a label from a fixed list (missing_code, code_not_string, path_not_string, files_not_array, unknown_language or backend_400). These labels do not include the submitted code, the file path or the error message. Each event also carries two labels from fixed lists: a traffic class (bot, internal, client or unknown) and a client family (for example claude, cursor, fixter or other), which the server works out from the User-Agent header and the client name. On a later request that has no client name of its own, the server uses the client name carried in the session ID; the User-Agent itself is not sent. Each event also carries name_family (the family matched by the client name alone, or named_unmatched when that name matches no rule, or none when there is no name), ua_family_raw (the family matched by the User-Agent alone, using only its first 512 characters) and classifier_version (the fixed label 2026-11-02.1). These are labels from fixed lists. The User-Agent text and any library version inside it are not sent. A calibration header, when present, is deleted from the request right after it is read, at the start of the request, so its value is not written to any log. The usage event carries only a boolean named calibration. The header value is not sent. Each event also carries surface, a label from a fixed list (default or core), that records which URL was used: default for /api/mcp-protocol and core for /api/mcp-core. That label is the URL, not proof of which product called the server. Events sent before this property was added do not carry it. Since 2026-10-28 each event may also carry ref, a short directory label taken from the request query when it is present and valid; it is not sent when missing or invalid. On initialize the server returns a random session ID (Mcp-Session-Id header) that also carries the client name and version; when the client sends it back, the events from that connection use that random ID, so they can be counted together. Without it, each request uses a new random ID. A tool call and its error event share one random request ID. The session and request IDs are not derived from your IP address, headers or credentials. Weekly anonymous client ID. When we turn it on, the PostHog event ID (distinct_id) is instead a keyed hash (HMAC-SHA256, with a secret kept on our server) of the client name, the client version, the network part of the address the request came from (the first three blocks of an IPv4 address, /24, or the first 48 bits of an IPv6 address) and the current ISO week. It lets us count roughly how many different clients use the server each week. The address and its network part are used only to compute the hash in memory; they are never stored, logged or sent to PostHog, and without the secret the hash cannot be turned back into an address. The ID changes every week, so it does not follow a client from one week to the next. It carries no IP address, no code, no headers and no tool arguments, and PostHog's IP geolocation is turned off for it. These events are sent from our server, so PostHog sees the server's address, not yours. PostHog keeps these events under its own retention terms.
Hosting. Vercel hosts hefestoai.narapallc.com and receives your IP address and request headers on every page load. Vercel's privacy policy applies to that processing.
Vercel Web Analytics. It loads on every page without a consent prompt and does not use cookies. Vercel records the page URL, the referrer, filtered query parameters, an approximate location (country, region, city), the operating system, the browser and the device type. Vercel identifies visitors by a hash of the request and discards it after 24 hours.
PostHog web analytics and session replay. Every HTML page loads posthog-js: the home page (including the pricing section and the buttons that open Stripe checkout), the contact page, the terms page and this page. Events go to PostHog US Cloud (https://us.i.posthog.com). The script sends a page view on each load, autocaptured clicks, and a session replay. It does not send the text you type into inputs. In the replay every input is masked, and the Stripe buy button, every iframe, and card number, expiry and security-code fields are replaced with an empty block, so payment fields are not recorded. Request and response bodies are not recorded, and the replay drops query strings from page URLs. The script stores an anonymous id in localStorage. It does not set a cookie and it does not identify you, so it does not create a PostHog person profile. Our PostHog project discards the client IP address, and each event asks PostHog not to look up a location from it. If the browser's Do Not Track setting is on, the script sends nothing. PostHog keeps these events and replays under its own retention terms. Session replays are kept for 30 days.
Microsoft Clarity. We use Microsoft Clarity to understand how visitors interact with our website. Clarity collects behavioral data such as clicks, scroll depth and session replays, and it uses cookies. It only loads if you accept the consent banner.
Google Fonts and Stripe scripts. The homepage loads fonts from fonts.googleapis.com and the Stripe buy button script from js.stripe.com, so Google and Stripe receive your IP address and browser details when the page loads.
AI crawlers. When the user agent matches a known AI or search crawler, our hosting logs record the company, the matched crawler token, the path, the time and the user agent (cut to 200 characters). Those lines contain no IP address and are kept for 1 hour. When PostHog is enabled, we also send one bot_visit event with only the company, the matched crawler token and the path. That event does not include the user agent or the IP address. We tell PostHog not to look up a location from it, and the event id is the crawler token (the text "bot:" plus the token), not a person. The event is sent from our hosting, so PostHog sees our server's address, not yours. PostHog keeps it under its own retention terms. These visits are not written to our database.
Agent card. Each request to /.well-known/agent-card.json stores the User-Agent (cut to 200 characters), the referring site (scheme and host only, such as https://example.com) and the time in our database (Neon Postgres). We do not store your IP address for this. Records created before this change also hold the IP address and the full Referer; a planned database change deletes those IP addresses and shortens those Referers. These rows are not deleted automatically. Ask us to delete them at support@narapallc.com. Once a day we send PostHog (US Cloud) the number of visits per day and per client name taken from the User-Agent (for example GPTBot or curl; other web browsers are counted as browser). The full User-Agent, the referring site and the IP address are not sent.
When you first visit our site, a consent banner appears at the bottom of the home page. Clarity only loads if you click "Accept." If you click "Decline," no analytics cookies are set. Your preference is stored in your browser's localStorage under the key "hefesto_cookie_consent" โ this is not a cookie and is not transmitted to any server. PostHog, described above, is separate from that banner: it loads on every page, does not use a cookie, and stops when Do Not Track is on.
If you join the email list, we store your email address, the name you enter (optional), your language and your subscription status in our database, and we send the emails through Resend. Our logs do not contain your email address; an error message shows at most a masked form such as a***@example.com, and runtime logs are kept for 1 hour. The unsubscribe link in every email stops the emails and keeps the record marked as unsubscribed. To have the record deleted, email support@narapallc.com.
Stripe processes payments. We do not store credit card information, and Stripe's privacy policy applies to payment transactions. After a purchase we store your email address, your Stripe customer, subscription and price IDs, the plan, the license status and expiry, and a keyed hash and the first 9 characters of your license key (for example HFST-AB12), which support uses to find your record. The keyed hash is an HMAC-SHA256 made with a secret that is kept outside the database. We cannot recover the full key; if you lose it, we issue a new one. The full license key is sent to you by email through Resend. During the move to the keyed hash we also keep the earlier unkeyed SHA-256 hash of the key; a planned database change removes it, together with the old column for plaintext keys. After that change, the database alone is not enough to check whether a key is valid. When the Pro CLI checks a license, it sends the key to /api/validate-license, which compares its keyed hash with our records. Current CLI versions send the key in the request URL, so it appears in our hosting provider's request logs, which are kept for 1 hour.
You can create an account at /signup with a name, an email address and a password. Accounts are managed by Neon Auth, the authentication service of Neon, our database provider, and are stored in our Neon database on AWS in the United States (us-east-1). Neon Auth stores your name, your email address, whether the email address is verified, a hash of your password (never the password itself) and, for each session, when it expires, the IP address it was created from and the browser user agent. To verify your email address, Neon Auth sends a numeric code to it through Resend, our email provider, and stores a verification record for the code. The code expires after 15 minutes. You enter it on the sign-up page, where you can also ask for a new code. Sign-up, sign-in and verification go through our server, which does not forward your IP address, so the IP address stored with a session is our server's. Our server limits sign-ups, sign-ins and requests for a new code, together, to at most 10 per minute and 100 per hour from one IP address (for an IPv6 address, its first 64 bits). Every attempt counts, including refused ones. The shared counters are kept in Upstash Redis under keys made from a SHA-256 hash of a fixed prefix plus the address, and the current minute or hour. The per-minute counter expires after 120 seconds and the hourly counter after 2 hours. Anyone who knows an IP address can compute the same hash, so we treat the hash as personal data. If Redis cannot be reached, the server counts in its own memory, keyed by the address itself. Those counters are never written to disk, they are cleared when the server instance restarts, and the oldest entries are dropped when they hold 10,000 addresses. When a limit is reached, our server logs one line that says so and which limit it was, with no IP address and no email address. A session cookie keeps you signed in until you sign out or the session expires. The email address shown after you sign in is excluded from session recordings. To delete your account, email support@narapallc.com from the address you signed up with.
The CLI and the GitHub Action analyze your code where they run and do not upload it. The MCP analyze tool is different: it runs on Narapa's server on the files you send and stores nothing. This section describes hefesto-ai 4.18.0 and later, checked against the source of the current release, 4.22.0.
Where usage events go. The CLI sends usage events by default to PostHog US Cloud (https://us.i.posthog.com), which stores them on servers in the United States. Our PostHog project has the setting "Discard client IP data" turned on, so PostHog does not store your IP address with the events, and every event asks PostHog not to look up a location from it. Every event is also sent without a PostHog person profile.
Install ID. Each event carries a random ID (a version 4 UUID) that the CLI creates once and keeps in ~/.hefesto/telemetry_state.json. It is not derived from your name, account, machine or repository, and we do not link it to them. It does let us count runs that come from the same installation. Delete that file to get a new ID.
The first run. The first run that could send an event prints a notice to stderr and sends nothing. Events start on a later run.
Events and fields. There are five events. Every event carries the event name, the install ID and a source: cli (you ran the command), hook (a git hook), ci (CI=true) or action (GitHub Actions).
| Event | When | Extra fields |
|---|---|---|
| first_run | Once per installation, on the first run after the notice | None |
| analyze | After each hefesto analyze | Hefesto version, OS (for example linux), Python version (major.minor), environment labels (ci, github_actions, gitlab_ci, jenkins, circleci, docker, kubernetes, dogfood, or user when none apply), install type (pypi or editable), and the counts: files analyzed, issues found, duration in milliseconds and exit code |
| hooks_installed | hefesto install-hooks installed at least one git hook | None |
| hook_blocked | A git hook run of hefesto analyze failed the gate | None |
| upgrade_cta_shown | The CLI printed a notice about a paid feature or a usage milestone (only in an interactive terminal, never in CI) | Which notice it was (for example gated_scope or analyze_milestone) |
The events never contain source code, file paths, file names, findings, repository names or your IP address.
Turning usage events off. Any one of these stops every event:
Version check. After hefesto analyze, the CLI asks pypi.org for the latest hefesto-ai version, at most once every 24 hours (the answer is cached in ~/.hefesto/.latest_version). The request sends no Hefesto data, but pypi.org sees your IP address, like any web request. From version 4.20.3, the CLI skips the check, with no request and no cache write, when HEFESTO_NO_UPDATE_CHECK=1, HEFESTO_TELEMETRY=0 or DO_NOT_TRACK=1 is set; hefesto telemetry disable does not skip it. In version 4.20.2 and earlier, nothing turns this check off, including the settings above.
GitHub Action. The Action's telemetry input defaults to 0, and with that default it sends no usage events. With telemetry: 1 (or true), the CLI inside the Action sends the events above with source action, and the Action also sends one request per run to https://hefestoai.narapallc.com/api/telemetry with the Hefesto version and the file count, issue count and exit code. We store that request in our database (Neon) without your IP address, and copy it to PostHog as described under Older CLI versions. Those rows are not deleted automatically today.
Older CLI versions. Versions before 4.18.0 sent their usage pings to https://hefestoai.narapallc.com/api/telemetry instead of PostHog. We store those pings in our database without your IP address, and they are not deleted automatically today. Once a day we copy these pings and the GitHub Action pings to PostHog (US Cloud) for analysis: the time, the event name, the Hefesto and Python versions, the operating system, the file and issue counts, the duration, the exit code, the environment labels (such as ci or docker), the install source and the random ID that those versions created once and kept in ~/.hefesto/.session_id. A text field that does not have the format the CLI sends is replaced with other or left empty. No IP address is copied. Each copy replaces the previous one, so a ping deleted from our database also leaves PostHog at the next copy.
| Recipient | What it receives | How long it is kept |
|---|---|---|
| Vercel (hosting) | Every request, including MCP and analyze requests with the code you send; IP address and headers; runtime logs; Web Analytics | Runtime logs: 1 hour on our plan. Web Analytics visitor hash: 24 hours |
| Upstash (Redis) | Rate-limit counters for the MCP analyze tool and for account sign-ups, sign-ins and new codes, keyed by a SHA-256 hash of your IP address | 120 seconds; 2 hours for the hourly account counter |
| Neon (Postgres) | Email list, licenses, accounts (Neon Auth: name, email address, password hash, email verification records, sessions with IP address and user agent), GitHub Action pings sent with telemetry: 1 and usage pings from CLI versions before 4.18.0 (no IP address), agent card visits (User-Agent, referring site) | No automatic deletion; delete on request |
| PostHog (US Cloud) | CLI usage events with a random install ID, unless you turn them off; the client IP address is discarded. Aggregate MCP event counts with a random session ID or a weekly anonymous client ID (a keyed hash; no IP address is sent), when enabled. Website page views, autocaptured clicks and session replays (inputs masked, payment fields not recorded, no request bodies, anonymous id in localStorage, no person profile). AI crawler visits with the company, the crawler token and the path, when enabled (no IP address and no user agent). A daily copy of the usage pings stored in Neon (older CLI versions and GitHub Action pings, with their random ID; no IP address) and daily counts of agent card visits per client name (no IP address, no full user agent, no referring site). No code | PostHog's retention terms for our plan. Session replays: 30 days. Copies from our database are replaced every day |
| pypi.org (Python Software Foundation) | The CLI's version check request (your IP address and a generic Python User-Agent; no Hefesto data) | PyPI's retention terms |
| Microsoft Clarity | Behavioral data, only after you accept | Microsoft's retention terms |
| Resend | Your email address and the email we send you | Resend's retention terms |
| Stripe | Payment details at checkout; browser details when the buy button script loads | Stripe's retention terms |
| Google Fonts | IP address and browser details when the fonts load | Google's retention terms |
Under GDPR, you have the right to access, correct or delete the data we hold about you. We can find and delete email list records, license records and older agent card rows that still hold an IP address. Newer agent card rows contain no data we can link to you. We cannot link hashed rate-limit keys or expired logs back to you. To exercise your rights or ask questions, contact us at support@narapallc.com.
We may update this policy. Changes will be posted on this page with an updated date.