LetterScript publishes a connector — a Model Context Protocol (MCP) server — so the AI assistant you already work in can read the transactions you choose. Ask Claude for the inspection contingency date in the middle of drafting an email to your client, and it answers from your actual file instead of asking you to paste the contract.
The connector address is https://mcp.letterscript.com/mcp. Everything below is one-time setup per assistant.
Availability: in Claude you can find LetterScript in the connector directory and add it without typing an address. Every other assistant — ChatGPT, Gemini, Perplexity and the rest — still needs the address pasted in by hand.
What the assistant can read
You approve a specific list of transactions, and the assistant can't read any transaction outside that list. Here's what it can read:
- Your transactions — status, address, stage, deal type, headline price, projected close of escrow.
- The transaction record — the reconciled facts (buyer, seller, prices, deposits, key dates, escrow and title), which document and page each was read from, cross-document conflicts, the party roster, and the fields a person confirmed by hand. Same record the in-app transaction assistant is grounded on.
- Deadlines — live contingency, deposit, disclosure and close-of-escrow dates, how each was derived, and the suggested next step. See Deadlines and the timeline.
- Point-of-sale requirements — what the property triggers, with legal citations, plus the hazard, fire, flood, septic and special-tax facts on record. See Point-of-sale requirements.
- The disclosure checklist — which requirements your uploads already cover and which are still missing. See The POS disclosure checklist.
- The document list — titles, processing status, classified type, the forms found inside each file, effective dates and folders.
- Search inside those documents — short passages with the document and page each came from.
- Your usage — transaction balance and the allowances on your free transaction. See Plans and transactions.
- Your account email address — only if you approve it. An assistant may also ask for the address you sign in to LetterScript with, and whether we've confirmed it's yours. It's given out as your account's address only when the approval screen lists Your email address and you approve. See Your email address below.
What it can't do
The connector is read-only. There is no write path in it at all — an assistant connected to LetterScript cannot:
- Send anything. No email from your transaction inbox, no invitations, no replies.
- Change anything. No edits to the transaction record, no new or moved deadlines, no uploads, no deletions, no collaborator changes.
- Spend anything. It can't buy transactions or touch your card.
And some things simply aren't in what it returns, whatever you ask:
- Taxpayer identification numbers (Social Security and Employer Identification Numbers) are removed from the transaction record entirely, and masked in document text. Detection on a scanned document is not perfect, so treat it as a strong filter rather than a guarantee for a file that arrived as a scan.
- Wire and banking instructions — routing, account, IBAN and SWIFT numbers, and whole paragraphs about moving money — are replaced with a redaction marker that points you back into the app.
- Party phone numbers and email addresses are not in the party list or the transaction record the assistant reads. They can still appear inside a passage that a document search returns — document text is redacted for taxpayer identifiers and banking instructions, not for contact details.
- Whole documents are not readable. Search returns short passages, never the file, and no download link ever leaves the app.
- Transactions you didn't approve, and anything in a workspace you don't have access to.
Add it to Claude
From the directory (easiest): open Settings → Connectors, search for LetterScript, and add it. The listing carries a Community label — how Claude marks a connector built by someone other than Anthropic. Anthropic screened it before listing it but has not reviewed it in depth, so Claude shows a reminder to that effect before you connect.
By address instead: Settings → Connectors → Add custom connector, paste https://mcp.letterscript.com/mcp, and add it. Or use this prefilled link, which opens the same dialog with the name and address already filled in: Add LetterScript to Claude.
In Claude Desktop: the same Settings → Connectors → Add custom connector flow. A connector you added on the web shows up in Desktop automatically when you're signed in to the same account, so you only need to do this once.
If you are shown an authentication choice, pick Sign in when needed. Adding by address, Claude checks the server and pre-selects No sign-in — because LetterScript lists what it can do before asking who you are. Change it to Sign in when needed. If you leave it on No sign-in, every question comes back as an authorisation error and you're never shown the sign-in screen. Leave the OAuth client on the recommended Use Claude's published identity.
It's worth getting right first time: the setting can't be changed afterwards, so fixing it means removing the connector and adding it again.
Claude won't ask you to sign in while you're adding it — that happens the first time you actually ask something about a transaction. That part is expected.
On a Claude Team plan? Ask your admin
On Team and Enterprise plans, only an Owner or admin can add a connector for the organization. If you're a member, you'll see a Request button rather than an Add button. Use it, or send your workspace owner this article and the address https://mcp.letterscript.com/mcp — once they add it, it appears for everyone and each person still signs in to their own LetterScript account and approves their own transactions. Nobody inherits anyone else's access.
You can also add it in a personal Claude account in the meantime; the two are independent.
Add it to Claude Code
One command, from any directory:
claude mcp add --transport http letterscript https://mcp.letterscript.com/mcp
Then run /mcp inside Claude Code and pick letterscript to sign in. Your browser opens the LetterScript approval screen and hands the session back to the terminal when you approve.
Add it to ChatGPT
Until OpenAI lists it, ChatGPT reaches the connector through developer mode: Settings → Connectors → Advanced → Developer mode, then create a connector, paste https://mcp.letterscript.com/mcp, and choose OAuth authentication. ChatGPT scans the tools, and asks you to sign in on your first real question.
Caveat: developer mode isn't on every ChatGPT plan, and which plans have it has moved more than once. If the Advanced section doesn't offer it, your plan doesn't have it yet — there's nothing to enable on the LetterScript side.
Gemini Spark and other paste-URL assistants
Google's Gemini Spark custom apps accept an MCP address directly: create a custom app and paste https://mcp.letterscript.com/mcp. Google's own eligibility rules apply — personal Google accounts only, United States, 18 or over, on Google AI Pro or Ultra. Work and school accounts aren't eligible.
Perplexity, Grok, Mistral Le Chat, and Notion custom agents take a pasted MCP address too. They register with LetterScript automatically, so there's no extra step here — but availability changes often, some tiers require a workspace admin to allow custom servers, and we can't verify every one of them. If the assistant asks for a "server URL" or "MCP endpoint", ours is the same address.
Approving access
The first question that needs your data sends you to a LetterScript approval screen. Read it before clicking Approve — it's the only place the scope of the connection is set.
- Check the hostnames. The screen shows the host the request came from and the host you'll be sent back to. If either is unfamiliar, deny it. A line reading "This app has not been verified by LetterScript" means exactly that: only continue if you started the connection yourself, moments ago.
- Pick the transactions. One transaction is selected by default — the most recent. Tick only what that assistant needs. All current and future transactions is also there, and it asks you to prove it's you before it will go through — because it hands over every transaction you open from then on. Which proof depends on your account: your LetterScript password if you have one, or a six-digit code we email you if you sign in with Google, Microsoft or Apple. That code works only for the request in front of you, and it expires in a few minutes. The email contains no links — we never ask you to click one to approve access. If you get a message that does, it isn't from us.
- Check whether it asks for your email address. If the list of permissions includes Your email address, approving also shares the address you sign in with and whether we've confirmed it's yours. It's part of the same approval — you can't approve the transactions and leave the email address out. If you don't want to share it, choose Deny; the assistant gets nothing at all.
- Note what it costs. Searches through the connector draw on that transaction's AI processing allowance, with their own daily and lifetime ceilings on top, so a chatty assistant can't drain a file. Reading the record, deadlines, checklist or document list costs nothing.
Approve once per assistant. You'll get a confirmation email naming what was connected, with a disconnect link — if that email ever arrives and it wasn't you, use the link immediately.
Your email address
An assistant may ask which LetterScript account it's connected to — for example, so an organization that manages its own assistant workspace can check you're using a work address. Your email address is only ever part of that answer if the approval screen listed Your email address and you approved it; if that line wasn't on the screen, the answer never includes it. Like any other contact detail, your address can still turn up inside a document or email passage that a search returns — a purchase agreement's broker section, or the header of a message filed to the transaction.
- What it can get. At most: a stable account identifier — a random ID that stays the same for your account, not your name — plus, only if you approved Your email address, the address you sign in with and whether we've confirmed it's yours. Nothing else about your account comes back this way: no name, no phone number, no workspace, and no transactions.
- "Confirmed" means proven. We report the address as confirmed only after you've shown you control it: you signed in with Google; you signed in with Apple and Apple confirmed the address; you completed a password reset (we email a code to the address and you type it back); or, on the approval screen, you entered the six-digit code we emailed you. Creating an account with a password doesn't confirm the address on its own — we don't send a confirmation email at sign-up — and signing in with Microsoft doesn't count as proof, because the sign-in name Microsoft gives us isn't guaranteed to be an email address you control. Until one of those has happened on your account, the assistant is told the address is unconfirmed, and an assistant that requires a confirmed address may turn the connection away. Proof counts from when we started recording it, not retroactively: if you use Google or Apple and an assistant reports your address as unconfirmed, sign out of LetterScript and sign back in with Google or Apple. If you use a password, resetting it with Forgot password? on the sign-in page confirms it.
- It's recorded, without the address. Each time an assistant asks, our service logs record the time, which connection asked, and the assistant's host — never the address itself. It isn't written to your Audit log or recorded against any transaction.
- It stops when access stops. Disconnect, or turning off Allow AI assistants to access this workspace, ends it immediately along with everything else the connection could read.
Three things to ask it
- What are the upcoming deadlines on my Elm Street transaction?
- Which point-of-sale disclosures does this property need?
- What does the purchase agreement say about the inspection contingency?
Name the property the way you'd say it out loud — the assistant lists your transactions first and matches on the address.
Disconnect
Either end works, and either end is immediate.
From LetterScript: Settings → Integrations → Connected apps (under AI assistants at the top of that page — the #connected-apps link in your confirmation email jumps straight to it) lists every assistant you've approved, with what it can read, which transactions, and when it last called. Disconnect cuts it off — the next call it makes fails, and it can't renew itself.
From the assistant: remove the connector in Claude, ChatGPT or Gemini. That stops it calling us, and you can still disconnect the grant from Connected apps afterwards to close it out on our side too.
By resetting your password: resetting it with Forgot password? signs out every browser and app session on your account and disconnects every assistant connected to it — each assistant's next call fails and it has to be approved again, a connection someone was partway through approving doesn't go through, and anyone who was signed in has to sign in again with the new password. Forgot password? works for accounts that sign in with a password; if an account using your address was set up with Google, Microsoft or Apple sign-in instead, contact support.
If a connection is listed as Reconnect required, we've changed what the approval screen asks for since you approved it. Nothing is broken and nothing has leaked — disconnect it and add it again to re-approve.
Turn it off for the whole workspace
Under Settings → Integrations, Allow AI assistants to access this workspace is the master switch. Off, every connected assistant stops getting data immediately, including ones already approved, and no new connection can be approved. Your existing grants are kept, so switching it back on restores them.
Security posture
- Read-only by construction. Every tool the connector publishes is a read. There is no write path for an assistant to find.
- Scoped to what you approved. Access is granted per assistant and per transaction — plus your account email address, only if the approval screen listed it and you approved — and expires unless the assistant keeps renewing it. Disconnecting kills the renewal too.
- Redacted before it leaves. Taxpayer identification numbers are dropped from the transaction record outright. In document text — including titles and filenames — taxpayer identifiers and wire or banking instructions are masked before anything is returned; that part is pattern matching, so a scan that garbles the digits can defeat it. Document text is also fenced and labelled as untrusted content, so wording hidden inside a PDF somebody sent you reaches the assistant as data to report, not as an instruction addressed to it.
- Recorded. Every call is logged in our service logs, with the time and the network address it came from — never your question, your conversation or document text. Searches that run over a transaction's documents are also recorded in the Audit log of the workspace that owns the transaction (Settings → Audit log, which you can filter to one transaction), with who ran the search, the assistant, the transaction, and how much was returned — never the text of your question. Approving or disconnecting an assistant is recorded in your Audit log too. If an assistant asks which account it's connected to, our service logs record that as well — the time, the connection and the assistant's host, never your email address — but it isn't written to your Audit log or recorded against any transaction.
- We don't train on it. LetterScript doesn't use your documents to train AI models, and neither does the service that reads them for us. See Security & privacy.
- What the assistant does with it is the assistant's policy, not ours. Once an answer reaches Claude, ChatGPT or Gemini, that provider stores and retains it under its own terms, on its own systems. We can't control or delete it there. That's the honest trade of connecting anything, and it's the reason the transaction picker defaults to one transaction rather than all of them.
One standing rule a connector doesn't change: LetterScript will never ask you to move funds or share an access code because an assistant told you to. Confirm wire instructions by phone, on a number you already had.
LetterScript isn't affiliated with, endorsed by, or partnered with Anthropic, OpenAI or Google — product names here belong to their owners, and we support their assistants the same way we'd support any other.