Your files stay on your Mac, unless you let Claude read them.
Written against how Lyndran is actually built, not copied from a template. Where something is unfinished, it says so.
- Your files are never uploaded to us. Reading them — the text out of a PDF, a Word file, a spreadsheet, and a scan read with your eyes’ worth of accuracy — happens on your own machine, in the app itself. When you ask a question about what was read, the part it has to think about goes to the model for that one thought, and the app says so as it happens, with a count of the words going. Exactly how much, and when, is below.
- With Claude connected, Claude reads only what you switch on. Every kind of reading starts off; each one shows on your Mac with how many words went. What Claude reads goes to Anthropic under your own Claude account, not through us, and every change is still a card you press. How that works.
- Lyndran can only see six folders, plus any you add yourself. Documents, Downloads, Desktop, Pictures, Music, Movies — and folders you choose in Settings, by hand, up to twenty. Your home folder itself is excluded, on purpose, and no task can add a place.
- We keep no history of your tasks on our servers. Your tools, your results and your history live in a folder on your Mac that you can open in Finder. The three small exceptions, all things you switch on yourself, are named below.
- No advertising, no analytics SDKs, no third-party trackers, no cookies on this site or in the app. We do not sell or share anything about you, ever.
- “Forget everything about me” is one press, and it means it.
What stays on your Mac
Lyndran is a native macOS app. It is not a website wearing a window, and it is not a thin client for a server that does the real work. The work happens here.
Everything the product accumulates about you is stored in ordinary files on your own disk, inside the app’s support directory, and you can open it in Finder:
- The tools you build, which are the saved shape of a task that worked.
- The results of past tasks — the tables, ledgers and drafts Lyndran hands back.
- Your reminders and schedules, in a plain JSON store.
- Your preferences, including which tasks you have granted a standing “always allow.”
- A short digest of recent turns so follow-up questions like “open the third one” resolve. See below for what is in it — it is smaller than you would expect.
- A small log of Lyndran’s own steps: which step ran, whether it worked, and how long it took. It never holds file names, page contents, messages, email addresses or account details, and it stays under about 2 MB. Beside it, Lyndran notes its own rare internal errors; such a note can name a window or a file it was handling at that moment. Neither is ever sent anywhere, and Settings shows you where the log is.
None of this is synchronised to us. There is no account you can log into on a website to read your Lyndran history, because that history does not exist anywhere but your machine.
What leaves your Mac, and exactly when
Lyndran needs a language model to plan a task. That is the one moment anything travels, and it travels for the length of one request.
The reply is a plan: a short list of steps in a fixed format. Your Mac then runs those steps locally. Finding your files, reading the text out of a PDF, reading a scan, searching inside documents, taking the screenshot, doing the tidying — all of it happens on your machine.
Reading is the app’s own work in one more place than it used to be. A Word file, a spreadsheet or a LibreOffice document was until recently handed to macOS’s Spotlight importer, a system service, which read it and passed the text back. Lyndran now parses those bytes itself, inside its own process. Opening a document to read it involves no other program and no network at all.
We want to be exact about the last row of that table, because the older version of this page was not. “Find the invoice they sent in March and open it” sends nothing out of the invoice. “What are the payment terms on this contract?” has to send the part of the contract that answers it — there is no way to answer a question about a document without reading the document, and the reading has to reach whatever is doing the thinking. What is capped is how much, and how many things at once. Reading and searching a document are now local from end to end, which makes that exception sharper rather than softer: the only words of yours that travel are the ones answering a question you asked, and you now watch them being counted as they go.
If you switch on “Work step by step”, more of what Lyndran reads goes to the model, and it tells you each time. It is off unless you turn it on, in Settings. With it on, Lyndran looks at each result before choosing its next step, which is what makes it better at jobs with surprises. Looking means sending: the start and end of each reading, up to about 6,000 characters of it, goes to the model so it can decide what to do next, where otherwise only the part needed for a question you asked would go. Each time, the app shows how many words are going, at the moment they go. What it reads is still read on your Mac, your files are still never uploaded to us, and nothing it learns from a page, an email or a file can make it act without your press.
Sorting files into folders you name goes by their names first. When you ask Lyndran to sort a folder into folders of your own — “Customers, Suppliers and Money” — the names of the files in it go to the model, with your folder names, so it can say which file belongs where. Only for a file whose name doesn’t say does the start of what it contains go too: about the first 600 characters of each, never more than 24,000 characters for one sort, and never a file that looks like it holds a password, a key or a card number. Nothing moves until you have seen every move on a card and pressed it, and a file it could not place stays where it is.
The digest limit is deliberate and doubles as a security control. An unbounded label is how someone smuggles instructions into a model through a filename or a page title, so the length cap is enforced in code, not in a prompt. There is more on that on the security page.
Lyndran with Claude
Your AI tells you what to do. Lyndran does it on your Mac, and nothing changes until you say yes. When you connect Claude, Claude does the thinking and Lyndran does the reading and the doing on your Mac. This is what that means for your data.
Where it goes. What Claude reads goes to Anthropic, under your own Claude account and Anthropic’s terms for it, the same as anything else you share with Claude. With Claude on your Mac (the Claude app), it travels from Lyndran to the Claude app on the same Mac and on to Anthropic. It does not pass through Mainspring. Who you have connected and what you have switched on is kept in a file on your Mac.
Claude on the web, and ChatGPT. When you connect Claude on the web or ChatGPT to your Mac, they reach it through Mainspring’s relay. The connection is encrypted from them all the way to your Mac, so the relay passes it along without being able to read it. It sees only that a connection happened, when, from which service, and how much went through, and it keeps none of what was said.
Adding Lyndran from ChatGPT’s or Claude’s list of apps. Then they connect to one shared address run by Mainspring, mcp.link.lyndran.com, instead of your Mac’s own. Each request is opened there for a moment, to check it was allowed and to send it on to your Mac, and goes on to your Mac encrypted. Nothing in it is stored or logged: we keep counts and timings, never what was asked or read. You sign in on app.lyndran.com with Google, so we learn which account is yours and which of your Macs to ask, and nothing else. You allow it on your Mac, and pressing Disconnect there ends it at once.
No model of ours is used. While Claude is using Lyndran, Lyndran never calls its own model; that is a rule in the code, and it is tested. Nothing Claude asks or reads is kept by Mainspring, and only the shared address above ever opens it, to send it on.
Asking in Lyndran. With Claude connected, what you type into Lyndran’s ask panel is handed to Claude: it opens a new chat in your Claude app, on your Mac. It goes to Anthropic under your Claude account, never to Mainspring.
The 7 days free, and membership. To know whether Claude or ChatGPT may use Lyndran, your Mac asks Mainspring at most once an hour whether your account has a membership or is inside its 7 free days, sending your account’s sign-in token. Mainspring sees that a check happened, for which account and when — never what your AI asked or read. We store, for your account, when the free days end (set once, the first time Claude uses Lyndran, never extended), next to your membership from Lemon Squeezy. Your Mac keeps the last answer in that same file, so it keeps working without a connection for up to three days.
What never leaves, under any circumstance
- Your files themselves. No document, image, PDF or spreadsheet is ever uploaded — not to us, and not to the model. Only the passage needed to answer a question you asked about one, within the cap above.
- Anything at all, when the task does not need thinking. Finding, opening, renaming, tidying, converting, extracting, searching inside and reading are local from end to end.
- The contents of web pages it reads on your behalf, beyond the clipped labels described above.
- Screenshots. A screen capture is downscaled and handed to the local step that needs it.
- Text read off a scan. Reading the words out of a photographed or scanned page uses a reader that ships inside macOS and runs on your machine. The image does not go anywhere.
- Your emails, messages, calendar entries or reminders. These are read and written through macOS itself, on your machine.
- Your passwords, keys, keychain or browser profile. A login you save for a site goes into the macOS Keychain; the model only ever sees a name for it, never a value, and the value is substituted in Rust one step before it is typed.
The folder boundary
File access is limited, in Rust, to six roots: Documents, Downloads, Desktop, Pictures, Music and Movies. Your home folder itself is excluded, because it is where .ssh, .aws, browser profiles and keychains live. Paths are resolved to their real location before the check, so a symlink or a ../ cannot be used to walk out of a root.
You can add more places yourself, in Settings, one at a time, through the Mac’s own folder chooser — up to twenty, kept in a file only you can read, and gone the moment you remove one. Nothing else can put a folder on that list: there is no capability, no task and no plan step that adds a place, so the model has no route to widening its own map. Your home folder, anything named as holding credentials, every hidden folder and macOS itself are refused even when you pick them by hand.
There is one way past the map, and it is yours: a file you attach yourself, through the Mac’s own open dialog, can be read wherever it lives for the rest of that conversation. It cannot be written to, moved or renamed, and the permission ends with the conversation.
What Mainspring stores, and why
Mainspring is the company behind Lyndran. On our own servers we hold as little as we can while still being able to run the thing and bill for it honestly.
If you bring your own model key
Lyndran talks to the model provider directly from your Mac. Your asks do not pass through us at all, and we hold nothing but your account record, if you made one.
If you are on a membership
Requests are relayed through our API so that usage can be metered. In that case we record, per request: your account identifier, a timestamp, the model used, the number of tokens in and out, and the resulting cost. This is what a bill is made of.
We do not retain the text of your asks or the plans that came back after the request has been served. We do not build profiles, we do not train models on your data, and we do not sell or rent anything about you to anyone. There is no advertising business here to feed.
Three exceptions, and you turn each of them on
If you send an ask from the web or from your phone for the Mac to pick up later, that sentence has to wait somewhere until the Mac comes for it, and that somewhere is us. It is the sentence you typed, held until it is collected.
If you switch on the mirror that lets your phone see what the Mac is doing, we hold the Mac’s current state and the labels of its recent tasks — what was asked, how it ended — so your phone has something to show you. Not the results, not the files.
If you switch on answering from your phone, the full words of a card waiting on your Mac — what it will do, and to what — are held with us while that card waits, so your phone can show you exactly what you are approving. They go when the card is answered or lapses, and your answer is held only until the Mac collects it. Files never pass through us.
Your Mac in a browser. At app.lyndran.com you sign in with Google, press Connect to your Mac and then Allow on your Mac once, and the browser then shows that Mac’s Lyndran: its projects, the cards waiting for you, what has been done, and the names of files. Those pages pass through Mainspring’s server (mcp.link.lyndran.com) on their way between the browser and the Mac, opened there for a moment to be sent on, not stored — the same way a request from ChatGPT or Claude passes. A file’s contents never take that road, with the one exception below that you switch on. That browser stays connected until you sign out or end it on the Mac, where it is listed under Settings by name (“Safari on a Mac”).
Asking a project from the browser. If you switch on Answer questions from my browser in Lyndran’s Settings on your Mac (it is off until you do), you can ask one of your projects a question at app.lyndran.com. Your Mac reads that project’s files, changes nothing, and answers. The answer can quote what is in your files, and it passes through Mainspring’s server on its way to your browser, opened there for a moment to be sent on and not kept. Your Mac logs every file it read for each question, and lists the question in Activity. Only you can ask, from a browser you let in, and never about a project someone else shared with you.
Photos the AI looks at. When you ask for photos by something your Mac can’t recognise on its own, like damage or a paint colour, Lyndran asks you first on a card that names the folder and how many photos. Only if you press Yes do small copies of the newest 40 photos in that folder go to the AI to look at, once. If you switch on Let the AI look at photos without asking each time in Lyndran’s Settings (it is off until you do), the card is skipped. The copies are 512 pixels on their longest side, carry no location or camera details, and go the same way your asks do; Lyndran doesn’t keep them. Your original photos never leave your Mac, and Claude or ChatGPT connected to Lyndran can’t use this.
Photos people sent you in Messages. When you ask for the photos someone sent you in Messages, Lyndran reads, on your Mac, which pictures came in, who sent them and when. It never reads what anyone wrote, and never the photos you sent. macOS only allows this after you switch on Full Disk Access for Lyndran in System Settings; Lyndran never asks for it on its own. Copying the photos into a folder happens only after you press on the card. To answer you, Lyndran may send the photos’ names, who sent them and when to its model, as it does with anything else it reads; the photos themselves stay on your Mac. Claude or ChatGPT connected to Lyndran can’t use this.
A recording turned into words. When you ask for a voice note or a meeting recording as notes, the words are recognised on your Mac by macOS’s own speech recognition (it needs macOS 26 or later). The recording never leaves your Mac. Nothing is sent to Apple: if your Mac doesn’t already have the speech model for that language, Lyndran says so and tells you where in System Settings to add it, rather than fetching anything. The words it produces are then read like any other document you ask about, so they go to the model the same way.
How far it is between places. When you ask how far it is between places, the place names go to Google Maps in Lyndran’s own browser, which opens Google’s directions page and reads the miles and minutes off it. Only the postcodes or town names go into the page’s address — never words from an email, a document or another page. If Google asks you about cookies first, Lyndran stops and says so; it never answers that question for you.
Your signature on a PDF. When you ask Lyndran to sign a PDF, it draws the signature picture you name onto a new copy of the PDF, on this Mac. The picture never goes to the model; only its file name does. Every signing needs your press on the card, and it can never be set to “always allow”. The signed copy loses the original’s links and bookmarks, and the card says so.
Gmail labels. When you ask for emails to go into a Gmail label, Lyndran makes the label in your Gmail (in its own browser, signed in as you) if it is not there, then files the emails. It presses only Gmail’s own buttons, and only on mail.google.com.
People in your Contacts. When you ask for someone in your contacts, like “the electrician’s number”, Lyndran asks the Contacts app on your Mac for the people whose name, company, job or note has the words you used, and at most twenty of them. It never reads the whole address book, and with no words to look for it reads nothing. It never changes a contact. To answer you, Lyndran may send those people’s names, numbers and email addresses to its model, as it does with anything else it reads. Claude or ChatGPT connected to Lyndran can’t use this.
One project for one AI. You can connect ChatGPT or Claude to a single project instead of your whole Mac: on that project’s page, press Connect this project to ChatGPT (or Claude), sign in with Google and press Allow on your Mac. That connection reads only that project’s files, databases, notes and memory, and asks for changes inside it with the same cards; it cannot reach your mail, your screen, your browser or your other projects. What it saves to the project’s memory is kept as its own words, never as yours. It is listed on its own in Settings, and disconnecting ChatGPT or Claude there does not end it; it ends when you disconnect it, or when the project is archived. As with any connection, what your AI reads goes to OpenAI or Anthropic under your own account, through Mainspring’s server without being kept.
Asking Lyn in the browser. With a membership, what you type into the Ask Lyn box at app.lyndran.com goes from Mainspring’s server to Anthropic to be answered, with the last few turns of that conversation in that tab and nothing else: never your files, your Mac’s contents or a shared project’s material. It is metered in carrots. Mainspring passes it on and does not keep it; the conversation lives in that browser tab. When you press Do it on my Mac, the words are kept on our server until your Mac collects them or you withdraw them, and your Mac shows them as a card before anything changes.
Files between a browser and your Mac go directly. You link a browser for files by scanning a code on your Mac’s screen. That code carries a key that stays in the browser and on the Mac and is never sent to us. To start a transfer, the two sides need to find each other, so we pass a short introduction between them: the network addresses each one can be reached at, sealed with that key so the server passing it along can’t alter it. We keep it for at most 90 seconds, then it is gone. To learn its own outside address, each side also asks a public address server run by Google, which sees that address and nothing else. The file itself then goes straight from one to the other, encrypted, and never through us or Google. If the two can’t reach each other directly, nothing is sent, and Lyndran says so. We don’t relay files as a fallback.
An ask that arrives from your phone waits for you to start it, unless you have also switched on letting plain tasks start by themselves. Even then, anything that sends, buys, posts, deletes or touches money stops at a card for your press.
None of these is on unless you turn it on, and each is described where you turn it on rather than only here.
Server logs from our API hold ordinary operational data — IP address, timestamp, path, status code — for a short window, because that is how you find out why something broke. They are not analysed for anything else.
Lyndran on your phone
The phone app is a remote, not a second Lyndran. It plans nothing and reads none of your files, except the ones you pick for your Mac. It sends asks to your Mac, shows what the Mac is doing, lets you answer a card waiting there, stops a task, and moves files between the phone and the Mac.
Signing in. Either with Google, through Firebase Authentication as described below, or by scanning a code shown on your Mac. A code works once and for two minutes. When a phone signs in with a code, we record that device on your account: a random identifier, the name the phone gives itself (usually its make and model, such as “Google Pixel 8”), and when. Removing a device from your account today signs out every session on it, the Mac’s included, and each has to sign in again. We say that plainly because it is blunter than it should be.
Files between your phone and your Mac never pass through us. They go directly over your own Wi-Fi, encrypted, and only when both are on the same network. Your Mac only listens once you add a phone, only answers devices on your own network, and only a phone that scanned the code on its screen. The phone checks it is talking to that Mac by the fingerprint in the code. A file sent from the phone lands in Downloads, in a folder called “From your phone”, marked as downloaded, and Lyndran never opens it. A file sent to the phone waits for your press on the Mac first.
When your Mac asks your phone for something. Lyndran on your Mac can ask the phone for files or photos, for how the phone is doing (battery, free space, model, Android version), or for the text on its clipboard. The phone shows you exactly what would go, and nothing goes unless you press. You choose the files yourself in Android’s own picker; the app cannot look through your phone. What you send goes to your Mac over your Wi-Fi. To answer you, Lyndran on your Mac may then send what it reads to its model, exactly as it does with a file on the Mac.
What the phone keeps. The key that links it to your Mac is kept inside the app and left out of phone backups and device transfers. Uninstalling the app removes it. Forget the phone on your Mac, and the key stops working.
What the phone sends to us is only what the three exceptions above already describe: an ask you typed, the mirror of what your Mac is doing if you turned it on, and your answer to a card if you turned that on. A request to stop a task is also held for two minutes, until the Mac collects it.
This website
The page you are reading right now makes zero requests to any other host. No fonts from Google, no analytics, no tag manager, no embedded video, no A/B tooling, no session recorder. Every asset is served from this domain, and the page’s Content Security Policy forbids anything else.
We set no cookies. The only thing stored in your browser is a single localStorage entry remembering whether you chose light or dark. It never leaves your browser and it is not an identifier. Clearing site data removes it.
The site is hosted on Firebase Hosting, which keeps standard web server logs. The one request this site makes off-page is the waitlist form, described next, and only when you submit it.
The waitlist
This site no longer has a sign-up form. If you typed your email into the one it used to have — the one that said it would tell you when a Windows version lands — we still hold that address, the time you sent it, and which page it came from. We use it for exactly one thing: to write to you once, on the day there is something to tell you about.
We will not add you to a newsletter, sell the list, or hand it to an email marketing platform. Every message we send will carry a one-click unsubscribe. If you would rather come off the list now, write to us and we will delete the row — there is nothing else attached to it.
Accounts and sign-in
Signing in uses Firebase Authentication from Google. That means Google, as our processor, handles the credential itself: we receive a stable user identifier and the email address on the account, and we never see or store your password.
Your account record on our side is deliberately thin — an identifier, an email, when it was created, and your membership and carrot balance if you have one. You can use Lyndran without an account by bringing your own model key.
Who else touches it
We use a small number of suppliers, each for one job:
- Anthropic — the language model that plans tasks. It receives the material listed above and nothing else. Requests are sent through the commercial API, which is not used to train models.
- Google Cloud — Firebase Hosting for this website, Firebase Authentication for sign-in, Cloud Run for our API, Firestore for the ledger of accounts and usage, and Secret Manager for our own credentials.
- Google Play services (Android only) — the code scanner runs inside Google Play services on your phone, and Google’s scanner and sign-in libraries send Google their own diagnostics about how they are working. They do not receive your files or your asks.
- Google’s public STUN server (stun.l.google.com) — used only when you move files between a browser and your Mac. It sees the network address of the browser or Mac asking, and nothing else.
An earlier version of this page named Neon, a Postgres host, as the database. That was left over from an earlier plan and was never true of what shipped; the store is Firestore. We are correcting it here rather than quietly.
These are processors acting on our instructions. We do not have any other integrations, and we do not add one quietly — this list is part of the page and changes when the list changes.
Web searches. When you ask Lyndran to find something out in the world — a hotel near a job, a room to hire, a restaurant for a group — it searches Google in Lyndran’s own Chrome on your Mac, the same as typing the words into Google yourself. The search words go from your Mac to Google, under Google’s own terms, and the results are read on your Mac. A search is your own words, or a short line Lyndran writes after reading something of yours — “hotels near Leeds LS7” — at most 120 characters. Because the AI writes that line from what it read, it can repeat a few words of an email or a file, such as a name or a number; the raw text of an email or a file is never sent as the search. Nothing about a search passes through Mainspring. Opening a result is reading a page, and booking or paying anything still waits for your press. A link Lyndran opens is one you gave it, one it read on a page, or a site’s own search address filled in with your words (a flights search, say); a line it writes after reading something of yours is never opened as a link. A link read on a page can carry words in its address, chosen by whoever made that page.
Deleting everything
On your Mac
Press forget everything about me in the app. It erases the tools, results, history digest, reminders and preferences from disk. It is not a flag that hides them from the interface; the files go.
Dragging Lyndran to the Trash and deleting its support folder achieves the same thing by hand, and there is nothing left behind on the machine afterwards.
On our side
Write to us from the address on the account and ask for deletion. We remove the account record, the usage rows and any waitlist entry with that address. Because we do not hold your asks or your files, there is genuinely nothing else to remove. We will confirm when it is done, within 30 days and normally much sooner.
Devices you signed in with a code are removed with the account. On the phone, uninstalling Lyndran removes everything it kept.
Your rights
If you are in the UK, EU, or another place with comparable law, you have the right to ask what we hold about you, to have it corrected, to have it deleted, to get a copy in a portable form, and to object to processing. Ask, and we will do it — the honest answer is usually that the list is one row long.
Our lawful basis is straightforward: performing the contract when you use the product, and legitimate interests for keeping the service running and secure. The waitlist runs on your consent, which you can withdraw at any time.
Lyndran is not built for, or directed at, children under 16, and we do not knowingly hold data about them.
Changes, and how to reach us
If this policy changes in a way that matters, we will change the effective date at the top and say what moved. We will not quietly widen what we collect and leave you to notice.
Questions, deletion requests, or a correction to something on this page: ertansoftwaresolutions@gmail.com.
Lyndran is made by Mainspring. See also the terms of use and the security architecture.