Your client calls, without the names.
Hide every client name, company, email and phone number in a call before any AI reads it. A free app on your own computer does the hiding; your agent reads only the clean copy; the real names come back before you send anything.
If you've promised a client confidentiality, pasting their call into an AI app is a problem. This blueprint fixes it. Your agent installs Transcript Cleaner, a free app that runs on your own computer with no AI in it. You give the app a call transcript, and it swaps every name for a code: C001 for a client, O001 for their company. Your agent reads that clean copy, summarizes it or drafts your follow-up using the codes, and the app puts the real names back so you can send it. The original call and your client list stay in a private folder your agent never opens.
Transcript Cleaner on your computer, a clean-calls folder in your workbench that stays off GitHub, a rule that keeps your agent out of the private folder, and your first call cleaned and answered.
Client calls hold the best material for your agent: what they need, what they said, what you promised. Now your agent can work with all of it without a client's name ever leaving your computer.
The app finds names with an English name model and a few rules, on your computer. Two folders: a private one with the originals and the client list, and a clean one with the coded copies. Your agent only reads the clean one.
You click through the app: the tour, your name, which names to keep hidden, and the final OK on each call. Your agent downloads it, runs the setup, sets up the folders and the rule, and writes your replies.
Is this for you?
- Client calls, one-to-one or a handful of people. Made for this. Coaching, consulting, sales, service calls.
- You record in Fathom or Granola. Made for this too: the app can pull your calls straight in. Or you choose a transcript file, or paste one.
- Large group sessions, webinars, classes. Not designed for it. It works best with a few speakers.
- Calls in another language. Not yet. The name finder is English only.
Perks come with The Lab and get claimed from your dashboard: no second account, no second login. Transcript Cleaner pulls calls from both. Granola's connection needs its Business or Enterprise plan.
Two folders, and only one is for your agent
This is the whole idea. Once you see it, the rest is clicks.
- The private folder,
TranscriptCleaner-Vaultin your home folder. It holds your original calls and your client list, the only record of who C001 is. Only the app reads it. No agent, ever. - The clean folder,
library/calls/in your workbench. It holds the coded copies you approve. That is the only place your agent reads calls from. It stays off GitHub too: no names, but still a client conversation.
It writes the rule into your workbench's context/permissions.md, so both apps read it. In Claude Code it also adds a block for that folder to .claude/settings.local.json, then tries to open the folder and shows you that Claude Code refused. Codex has no per-folder block like that, so in Codex the written rule is what keeps it out, and your agent tells you so.
What the codes mean
- C001, C002... A person you saved as a client.
- O001, O002... A company you saved.
- C001-EMAIL, C001-PHONE. That client's email or phone number.
- EMAIL-1, PHONE-1. An email or phone number not saved to anyone.
- PERSON-1, COMPANY-1. A name the cleaner wasn't sure about, hidden until you review it.
A code means the same name in every call. Your agent gets a short note telling it to keep every code exactly as written and never guess who one is. The note comes from the app's own Put names back page.
The build, checkpoint by checkpoint
Nine checkpoints, and your agent waits for your yes at each one. Here are the main ones.
- It checks your computer. Windows 10 or 11, or a Mac, about 1.5 GB free, and the same computer your calls are on. It picks a folder for the app outside OneDrive and iCloud, because a syncing folder can break the setup.
- It downloads the app from its public page and runs its setup script. About 800 MB downloads once, into the app's own folder, in 5 to 15 minutes. Nothing system-wide, no admin password. You get a shortcut on your desktop.
- You take the tour. Guided tutorial walks you through one made-up call. Then Try sample calls in Settings gives you five more to click around in. Then you type your own name once, so it stays visible.
- It sets up the clean folder and the rule, and you point the app at it with Choose export folder.
- You clean your first real call. Add it, then on Review names decide each name it wasn't sure about: Save as a client, Fine to show or Keep hidden. Every answer has an Undo. Read the clean copy and click Approve clean copy and export.
- Your agent reads it and drafts the follow-up with the codes. You paste the draft into Put names back, read it once, and send it from your own email.
# Windows curl.exe -L -o "$env:TEMP\transcript-cleaner.zip" https://github.com/WMurley/transcript-cleaner/archive/refs/heads/main.zip Expand-Archive "$env:TEMP\transcript-cleaner.zip" -DestinationPath "<your Documents folder>" # then, inside the new transcript-cleaner-main folder: powershell -NoProfile -ExecutionPolicy Bypass -File app\setup-windows.ps1 -NoPrompt # Mac curl -L -o /tmp/transcript-cleaner.zip https://github.com/WMurley/transcript-cleaner/archive/refs/heads/main.zip ditto -x -k /tmp/transcript-cleaner.zip ~/Documents # then, inside the new transcript-cleaner-main folder: printf 'n\n' | bash "Setup Transcript Cleaner (Mac).command"
Your agent runs setup for you, so there are no security warnings to click through. If setup stops, it reads setup.log (which never holds a transcript or a client detail), fixes the cause and runs it again.
The app's public page has the steps for Windows and Mac: download the ZIP, move the folder into Documents, double-click the setup file. You will see one Windows or Apple warning on the way, and the page tells you exactly which button to click.
Pull calls from Fathom or Granola
Optional. Instead of choosing files, the app can bring your calls in directly.
- Fathom. In Fathom, Settings, then API Access, generate a key.
- Granola. Business or Enterprise plan. In the Granola desktop app, Settings, then Connectors, then API keys; the key starts with
grn_.
You paste the key into the box in the app's own Settings and click Connect. The app saves it in your computer's password store (Windows Credential Manager or the Mac Keychain) and never shows it again.
The key goes in one place: the app's own key box. Your agent never needs it, never asks for it, and never writes it into a file. If an agent ever asks you to paste a key, the answer is no.
What you can ask your agent, from now on
Five asks that work once a few calls are in library/calls/.
Read my newest clean call in library/calls/. Give me the three things the client asked for, what I promised, and any date we agreed. Keep every code exactly as written.
Draft my follow-up email for the newest call in library/calls/: thank them, recap what we agreed, and the next step. Keep every code exactly as written; I'll put the names back before I send it.
Read every clean call with C001 in library/calls/ and give me a one-page brief before my next call with them: what they want, what's open, and what I said I'd do.
Read all my clean calls from this month in library/calls/. What do clients keep asking about? Quote the line each time, with the call's file name.
Check my client-calls rule is in context/permissions.md and, in Claude Code, that the block for TranscriptCleaner-Vault is in .claude/settings.local.json. Don't open the folder to check. Just show me the two rules.
Words you now own
- Clean copy. A call with every name swapped for a code. The only version your agent reads.
- Private folder. Where the app keeps your original calls and the client list. Only the app opens it.
- Export folder. Where the app puts clean copies you approve. Here,
library/calls/. - Code. The stand-in for a name, like C001. The same name always gets the same code.
- Review names. The app's list of names it wasn't sure about. They're hidden until you decide.
- Put names back. The app's page that turns your agent's coded draft back into real names, ready to send.
Fair questions
No. The cleaning runs on your computer with no AI model. The app only goes online for setup, for Fathom or Granola when you choose, and when you click Check for updates.
Yes, sometimes. It catches most, and its final check points out anything that still looks like a name, but it's a safety net, not a guarantee. Read each clean copy before you approve it.
The cleaner wasn't sure whether it was a person. On Review names, click It's a tool. It remembers.
No. It has to run on the computer where your calls are, usually your laptop. Run this blueprint from Claude Code or Codex on that computer.
In the app: Settings, Updates, Check for updates, then Update now. Your calls, clients and settings stay as they are. To remove it, delete its folder and the desktop shortcut; your private folder stays until you delete it yourself.