Your browser, your keys

Who makes it

Sieve is a browser extension made by Grgur Damiani (di vergada, Zagreb, Croatia). Scoring runs on Jev, TypeSafe's decision model, reached through your OpenRouter key. It is an independent project and not made by TypeSafe. Short version of this page: the extension works in your browser with your own API keys and sends nothing to me unless you choose to: usage counts if you say yes to usage stats, your library if you turn on Your agent, and an email address if you join the waitlist on divergada.com/sieve.

What the extension reads

When you browse LinkedIn, X, Reddit or YouTube, Sieve reads what's shown for posts and videos that stay on your screen for about half a second: the author line and post text, or a video's title, channel and length and the text YouTube shows with it (description lines, chapter titles, YouTube's own summary). For a video that shows only its title, as on YouTube's home page and next to a playing video, Sieve asks YouTube for the start of that video's description. That request goes to YouTube only, without your YouTube cookies, so it isn't tied to your account or your watch history. You can turn it off in Sieve's settings. When it saves a LinkedIn post or writes a brief for one, it also reads that post's id from the page, so the saved post and the brief can link to the post itself. The id is not sent to TypeSafe or OpenRouter. On X, Sieve also reads each post's data from the page X has already loaded: the full text of a long post, which post it replies to, and the links to its pictures and its video. On a post's own page, that includes the author's other posts in the same thread. It asks X for nothing more. It does not read your messages, your profile, your connections, or any other page.

Where that text goes

Scoring goes to OpenRouter (openrouter.ai) with your OpenRouter key, and OpenRouter passes it to TypeSafe (typesafe.ai), which runs the Jev model, so both receive it. If OpenRouter can't reach Jev at that moment, the same request goes through OpenRouter to a fixed general model instead, and that model's provider receives it. If you saved only a TypeSafe key and no OpenRouter key, scoring goes straight to TypeSafe with your TypeSafe key. OpenRouter's, TypeSafe's and the model provider's own privacy policies apply. Nothing but scoring ever goes to TypeSafe. What is sent is the post text (for a YouTube video, its title, channel, length and the text described above, with at most the first 1,500 characters of its description), with the topics and description of yourself from your settings, so the score fits you. On Reddit, the facts you entered for Reddit go with it too, so the scorer can judge whether you could answer from your own experience.

OpenRouter (openrouter.ai), for everything else: only when you ask for something, text goes through OpenRouter to the model you chose, using your key. For a technique brief, it is the post's author line and text, with the topics and description of yourself from your settings. On X, it is the post's full text, and when the author's thread is on the page (as it is on a post's own page), each of their posts in it and the author and text of any post they quote. When the post or its thread has pictures, the links to up to ten of them go too, and the brief is written by the video model you chose, which can read pictures; the model's provider fetches the pictures from X through those links. The picture links are not stored. For a daily digest, it is the posts Sieve kept for you. OpenRouter's and the model provider's policies apply.

Watch it for me: only when you click it on a YouTube video, that public video's link, title and channel, with the topics and description of yourself from your settings, go through OpenRouter to the video model you chose (Google's Gemini by default). It watches the video and sends back a summary, and a technique brief when the video teaches one. Both are kept in your browser.

On X, Watch it for me works on a post with a video, again only when you click it: the link to the video file on X's servers, the post's account name and its text (at most 1,500 characters), with the topics and description of yourself from your settings, go through OpenRouter to the same video model, whose provider fetches the video from X through that link. The video link is not stored; the summary keeps the post's own link.

Nothing is sent to me or to any server I run, unless you turn on usage stats or Your agent (both below). There is no tracking and there are no ads.

Usage stats, only if you say yes

Sieve asks once whether it may send usage counts. It's off until you say yes. If you do, then once a day, for each day you used Sieve, it sends Google Analytics (Google Ireland Limited) a random ID made in your browser, the extension version, which scorer answered, whether Your agent is on, and counts: posts scored, saved, briefed, copied as a prompt, watched for you, put in a digest and exported, per platform. Nothing else: no post text, links, names, titles, facts, keys or anything you type. Google sees your IP address with the request, as with any request, and says Google Analytics doesn't store it. Google may process the data outside the EU. I keep it for 2 months and use it only to see whether people keep using Sieve. The legal basis is your consent.

Turn it off under Usage stats in Sieve's settings: sending stops, and the ID and any unsent counts are deleted from your browser. Settings shows a short install code while usage stats are on. If you tell me yours, I can ask Google to delete what it holds for your install.

Your agent, only if you turn it on

Sending your library to your coding agent is off by default and, for now, only for people I have invited. If you turn it on under Your agent in Sieve's settings, Chrome first asks you to allow two things: a sign-in window, and access to mcp.divergada.com. You then sign in through WorkOS (WorkOS, Inc.), with Google or with your email. WorkOS gives Sieve your email address, and Sieve makes a Sieve account for it. If you aren't invited, no account is made and Sieve asks WorkOS to delete the sign-in.

From then on Sieve sends your library to my server at mcp.divergada.com, which runs on Cloudflare (Cloudflare, Inc.). It sends the same items as Export library: saved posts with their text, summaries of videos watched for you, and your technique briefs. It does not send your digests, keys, settings or the facts you entered. It sends when you turn it on, when you press Send now, and within 15 minutes of a change, only if something changed.

Only coding agents you connect to your Sieve account, by signing in to them with the same account, can read your library. Besides the library, the server keeps your email address and how you signed in; a log of each upload (when, how many items, how big); a daily copy of your library for 7 days, so a broken upload can be undone; the results an agent records when you ask it to (a status and a short note); and when your agents last read and which items they opened each day, kept for 90 days.

Turn off deletes the Chrome copy from the server and signs this Chrome out. Delete my Sieve account deletes the account and everything the server keeps about it, including a library sent from Sieve for iPhone. Invitations run until 10 November 2026. If your access ends, the library, its copies, recorded results and agent activity are deleted 30 days later; your account and email stay until you delete the account or ask me to. Cloudflare and WorkOS may process this data outside the EU. The legal basis is your consent, and you can withdraw it at any time by turning it off or deleting the account.

What is stored, and where

Everything below is stored locally in your browser, in Chrome's extension storage (what Sieve's server keeps when Your agent is on is listed above):

• your API keys and settings,
• the facts you enter about yourself,
• usage counters (posts scored, briefs written, cost),
• posts that scored high and posts you asked for a brief on, kept for up to 30 days,
• summaries of videos watched for you and the technique briefs you asked for: the newest 300 of each are kept,
• the newest 60 daily digests you generated,
• if you turned on usage stats, a random install ID and the counts not yet sent,
• if you turned on Your agent, the sign-in for Sieve's server, a random device ID, and when your library was last sent.

Removing the extension deletes all of it. You can also reset the counters from the extension's popup.

Export library, on the digest page, saves a copy of what Sieve kept (posts, video summaries, briefs and digests) as one file on your computer. The file is made in your browser and sent nowhere.

What it never does

It never clicks, likes, votes, comments, posts, sends messages or scrolls on your behalf, and it never sells or shares data. It never installs or runs anything either: a technique brief is text, and it reaches a coding agent only if you copy it there yourself or, with Your agent on, when an agent you connected reads it from Sieve's server.

Source code

Sieve is open source under the MIT license, so all of the above can be checked: github.com/gdamiani1/sieve

The waitlist

The extension never asks for your address. The waitlist on divergada.com/sieve is the one place where you can give me one, and it is separate from the extension. It asks for your email address and nothing else.

If you sign up, your address is used only to email you about Sieve: the confirmation email, and then word when the Sieve account, the MCP server or Sieve for iPhone opens. Nothing counts until you click the link in the confirmation email. Every email has an unsubscribe link, and you can write to me at any time to have your address deleted. It is never sold, shared or used for anything else.

The list is kept by Infomaniak (Infomaniak Network SA, Geneva, Switzerland), which runs the sign-up and sends the emails, and its privacy policy applies to what it stores: your address, and when you signed up and confirmed. Infomaniak sends these emails through Amazon SES, Amazon Web Services' email service, so the emails pass through Amazon on their way to you.

When you press Sign up, your browser first asks Infomaniak for a small anti-spam puzzle and solves it. Like any request, that one shows Infomaniak your IP address. Your address and the puzzle's answer then go to Infomaniak through this site. Before you press Sign up, the page contacts no one: the Sieve pages on this site set no cookies and load no analytics or tracking script.

You sign up by choice, and you can take that back at any time with the unsubscribe link in every email. You can ask me for a copy of what is stored about you, to correct it, or to delete it. The list exists only for these launches: when they have opened, or when it's clear they won't, I send one last email about it and delete the list. If you think I've handled your data wrongly, you can complain to AZOP, Croatia's data protection agency (azop.hr).

Questions

Write to me directly: grgurdamiani@divergada.com