LEDGER LITE — INSTRUCTIONS FOR AI AGENTS ========================================= You are reading this because your user wants you to connect to Ledger Lite and work with their records. Ledger Lite is a free app that keeps a person's expenses, notes, forms, health records, and secrets (passwords and keys) on their own phone. It has no AI of its own — you are the AI. The app is the fixed home you write records to and read them back from. Your job with this document: walk your user through connecting you, one simple step at a time, then prove the connection works. Talk to your user in plain language. Give one step at a time, wait for them to finish it, and confirm before moving on. The taps happen on their phone, by their hand — you cannot tap for them, but you can make every step obvious. The whole setup takes about ten minutes. WHAT YOU NEED - Ledger Lite installed on the user's phone. Android: https://www.ledgerlitedwg.com — the Download APK button. iPhone (beta): https://testflight.apple.com/join/ThvGUNaX - Tailscale on the phone AND on the machine you run on, signed in with the SAME account on both (Step 1). - A caller key, which the user creates in the app and hands to you (Step 2). STEP 1 — TAILSCALE (the private link between you and the phone) Tailscale is a free app that links the user's own devices into a private network nobody else can join. You reach the phone through it — never over the open internet. 1. Ask the user to install Tailscale on their phone (Play Store or App Store), open it, sign in, and allow the VPN connection when asked. Tell them to leave it switched on. 2. Make sure Tailscale is installed and signed in on YOUR machine with the same account the user used on the phone. If you cannot install things yourself, tell the user exactly what to get and where: https://tailscale.com/download 3. Ask the user to open Tailscale on the phone, tap the phone's own entry, and read you its address. It starts with 100. — for example 100.101.102.103. You need it in Step 3. If either device is missing from the device list, it is almost always signed in with a different account. Have them sign out and back in with the same one. STEP 2 — THE KEY (the user hands you permission) Nothing can reach the user's data until they create a key for you in the app. Talk them through it: 1. "Open Ledger Lite and tap 'Connect your agent'. You will find it on the Home screen, at the top of Settings, and in the Vault." 2. The app checks Tailscale on both devices and walks through anything missing. Then it asks them to name the agent. Suggest a short name in lowercase letters, numbers, and dashes — for example the name the user calls you. 3. "Choose what it may do" — four switches. Reading records is on by default; most agents need at least that. Adding, changing, and deleting records is separate — ask for it only if the user wants you filing for them. Vault secret access is separate again, and only for the specific secrets they pick. Tell the user to turn on only what you actually need. 4. They review the summary and confirm with fingerprint, face, or PIN. Keys expire after 90 days unless they change the date or choose no expiry. 5. The app shows two things: an address that looks like http://100.x.x.x:8080/api/v1 and the key itself. The key is shown ONCE. Ask the user to paste both to you. The app also has a copy button with plain-English setup instructions written for agents — the user can paste you those instead; they carry the same address. 6. Last tap: "Start agent access and leave it running. On Android, if the phone asks whether Ledger Lite may keep working in the background, say yes — that is what lets me reach the phone while the app is closed. On iPhone, agent access runs while the app is open, so the user should keep it open while I work." STEP 3 — CONNECT AND PROVE IT - Base address: http://:8080/api/v1 - Send the key as a bearer token (Authorization: Bearer ) on every request. - First test: introduce yourself to the phone and read the user's expense categories. A working connection answers with your agent name and their real category list. Read the categories back to the user — that is their proof the setup worked. One exception: a key that can only write, or only fetch secrets, cannot run this test. That is on purpose, not a failure. - If the phone does not answer, check Tailscale first: it must be connected on BOTH the phone and your machine. That is the cause almost every time. Then check the user left agent access running in the app. GROUND RULES - The key is a secret. Store it the way you store credentials. Never print it back, and never put it anywhere the user did not ask for. - Stay inside the permissions the user picked. If a job needs more, ask them to widen your key in the app — do not work around it. - Everything you read is the user's private records. It stays between you and them. - The user can pause you, change what you may do, or replace your key at any time in the app. That is normal, not an error. PART 2 — WORKING WITH THE RECORDS (once you are connected) ========================================================= Setup got you connected. This half teaches the daily work: filing expenses, answering questions, keeping notes, and fetching forms, secrets, and health records. One idea runs through all of it: the app is the memory, you are the hands and the voice. Never work from your own memory of the records. Read them fresh, write carefully, and read back to prove it. Every example below is illustrative. The names, amounts, and dates are made up for teaching — no real person's records appear in this file. HOW TO WRITE ANYTHING (the rules for every kind of record) - Read before you write. Pull the current records first — the category list, the recent entries, the note you are about to change — so you know what already exists. - Check for duplicates before adding anything. A charge, note, or record that is already there must never be added a second time. - One record per real-world thing. One charge is one entry. Never merge two charges into one entry, and never split one charge into two. - Read back after writing and tell the user what the app now shows. - If a write's result is unclear — a timeout, an error after sending — read first before trying again. A short delay is fine; a doubled record is not. - If something could fit two places, ask. A short question beats a wrong guess the user has to clean up later. EXPENSES The most common job. The flow never changes: 1. The user hands you charges — a bank alert, a receipt, a list. 2. Read the expense categories. Use the user's own categories; if a charge fits none of them, ask instead of inventing one. 3. Check each charge for a duplicate across a wide window: same amount and a similar description, on or near the same date. 4. Show the user the table before you post anything — date, description, category, amount — and wait for their OK. 5. Post one entry per charge. 6. Read the entries back and confirm the count and the total. For example, three charges the user hands you might become: Date Description Category Amount 2026-11-03 Grocery store Food $62.40 2026-11-04 Fuel station Auto $38.10 2026-11-05 Streaming service Subscriptions $15.99 You show that table first, post after the OK, then read back: "Posted 3 entries, $116.49 total — the app shows all three." Not expenses: money moving between the user's own accounts — paying a credit card from a bank account, moving money to savings — is not spending. Never log a transfer as an expense. Answering "what did I spend on groceries last month?": query the records for that category and month, add up the actual entries, and answer with the total and how many entries it came from. Offer the list. Never estimate, and never answer from memory. If the records hold nothing for that period, say so plainly. BUDGET A budget in Ledger Lite lives on the categories: each category can carry a monthly target — the most the user plans to spend there in a month. Categories are grouped into sets, one set per book: the household, each property, a person, the business, the pets. How to work with it: - Read the targets together with the categories (Part 2, Expenses). - "How much is left for groceries?" — take the category's target, subtract this month's actual entries in it, and answer with both numbers: target, spent, left. From the records, never from memory. - When you file an expense, you can mention where the category now stands ("that puts Food at $1,240 of $2,000 for November"). If a charge would push a category over its target, say so before posting — then do what the user decides. The budget informs; the user rules. - Never change a target on your own. Changing the budget is the user's call, always. Drafting a starter budget (a common first job): read the user's last two or three months of entries, average each category, round to a clean number, and show the draft as a table — category, monthly target — for the user's OK before you save anything. A draft is a lever to start from, not a verdict on their spending. An example budget, edited from a real one (names generalized, amounts rounded) so you can see the shape: Household Food (groceries and eating out) $2,000 / month Amazon $2,000 / month Home maintenance $500 / month Entertainment $500 / month Church tithe $500 / month Gifts $250 / month Subscriptions (streaming, apps) $55 / month Miscellaneous $300 / month House 1 Mortgage $3,800 / month Utilities, internet, phone $425 / month Lawn $240 / month Pest control $50 / month House 2 Mortgage $3,400 / month HOA, utilities, internet $450 / month Cleaning $300 / month Lawn $200 / month Pest control $110 / month Cars and personal Second car $850 / month Gas and tolls $250 / month Auto insurance $200 / month Personal care $300 / month Gym $35 / month Dogs Vet plan $180 / month Grooming $300 / month Dog food $100 / month Business General $250 / month Fifty categories is normal for a real household. A new user's starter draft can begin with eight or ten and grow — the shape matters more than the count. PERSONAL AND BUSINESS, ONE HOME One Ledger Lite can hold a household and a business. Keep them apart with category sets — one set for the business, one for personal. File each charge to the set it belongs to, and questions answer per side: "what did the business spend this year" reads only the business set, and tax time is that set, already totaled. The one filing mistake that costs real time later is mixing the two. NOTES Notes hold the user's to-dos, lists, and project notes, many of them tagged by project. Tags are how notes split by purpose — one tag per project or topic. Find the user's list by its tag before you read or write; never start a second list for a purpose that already has one. - Read the relevant note before changing it. Add to it in place rather than starting a near-duplicate note. - Check an item off only when it is actually done, and leave a short note of the proof next to it — what you verified, and when. - When the user asks "what's on my list for this project?", read the note and answer from it. Do not rebuild the list from conversation memory; the note is the source of truth. Running to-do lists — the pattern that makes you useful across days: a note can be a living list you and the user share, usually one per project. Work it like this: - At the start of a job, read the list first. It tells you what is open, what is done, and what the user cared about last time — even if that was with a different agent, or weeks ago. Your continuity lives in their notes, not in your memory. - Work from the list: take the next open item, do it, check it off with its proof note, and move on. - When new work comes up in conversation, add it to the list in the same pass — an item that only lives in the chat is an item that gets lost. - Keep the list honest: never delete items (checked items are the history), never quietly drop one you find hard, and if an item is ambiguous, leave it open and ask about it in a batch with any others, as one short yes-or-no list. - End a work session by leaving the list current: what got done, what is next, what is waiting on the user. The next session — yours or another agent's — should be able to pick up from the note alone. FORMS Forms hold the ID numbers people always have to dig up: driver's license, passport, insurance, memberships. - Fetch exactly the field the task needs — no more. - Treat form contents like secrets: use them for the task at hand, do not repeat them back casually, and never copy them anywhere the user did not name. SCHEMAS — giving the user's information a shape Sooner or later the user starts tracking something of their own: the cars, the properties, the subscriptions, the dog's vet history, a collection. The thing that makes tracking work is a schema — the same small set of fields, filled in the same way, every time. You do not need a database for this; you need consistency. - When tracking starts, propose the fields first and show the user one filled example. Get their OK on the shape before you fill in twenty entries. - Then hold the shape. Every new entry gets the same fields in the same order. If a new fact fits no field, say so and extend the schema openly — never cram it into a field it does not belong in, and never quietly reshape entries that are already written. - Where a schema lives depends on the information: - Loose tracking (lists, logs, inventories) — a note, one entry per block or line, always in the schema's shape. - Official numbers and documents — a form. A form IS a schema the app holds for you: named fields, one record per thing. - Money — categories, with targets if the user budgets (see Budget above). Do not track spending in a note. For example, a note tracking subscriptions, one line per service, always in this shape: Name | Amount | Renews | Paid from | Used for Music app | $12 / month | 2026-12-01 | Joint checking | Family plan Cloud storage | $30 / year | 2027-03-15 | Business card | Backups And a form for a car might carry: plate, VIN, insurance policy number, insurer's phone, inspection due. With the shape fixed, you can answer "when is the inspection due?" or "what does streaming cost me a year?" by reading the entries — completely, every time, because nothing hides in a free-form paragraph. THE VAULT (secrets) - Fetch a secret by its exact name, only when a task needs it, and only if your key allows that secret. - Never print a secret back, never paste one into a chat, a form, or a document the user did not name, and never store one in your own notes or memory. - If a fetch fails, say the fetch failed. Never guess at, rebuild, or substitute a secret. HEALTH Steps, meals, and other health records live beside the money. - Log them exactly as the user gives them — their words, their numbers, the right date. - Never invent a reading. If the user did not give a number, there isn't one. WHEN SOMETHING GOES WRONG - Phone unreachable: it is Tailscale, almost every time (Part 1, Step 3). - Permission denied: your key does not cover that. Ask the user to widen the key in the app. Do not look for a way around it. - An answer you cannot support from the records: say what you checked and what is missing. A plain "I don't see it" beats a confident guess, every time. For the human version of this page: https://www.ledgerlitedwg.com/agent/ Guides with pictures, for the user: https://www.ledgerlitedwg.com/how-to/set-up-tailscale/ https://www.ledgerlitedwg.com/how-to/connect-your-agent/