Google Voice was free. This costs a little. Here's exactly how little, and what it bought me.

The bill

I pulled these numbers straight from Telnyx’s usage records while writing this. This is the last 30 days for all my lines, not just my main number:

The texting itself, the thing this whole story is about, was $2.48 for 332 messages. Add the fixed costs from the note under the table and the whole setup, AI receptionist included, comes to about $45 a month. That’s less than most cell phone plans.

My AI teammates, using the front door

Here’s a part of the payoff I didn’t expect. On September 28 my AI assistants got their own way into my texts: a small command-line tool that talks to my server (plus a version they can plug into directly) that can list unread threads, read a conversation, check calls and voicemails, and draft a reply.

Drafting is as far as it goes. Nothing gets sent until I approve the exact words and the exact person.

It also took some retraining. Twice, an assistant went looking for a text in my old Google apps out of habit. I told it the same thing both times:

That’s now a standing rule for every assistant I run. My texts live on my server, and that’s the only place anyone, human or AI, goes to read them.

Still on the list

It’s ErrorZap, so here’s what’s still broken. While writing this, I tried searching my texts through the assistants’ tool, and every search came back empty, even for common words. The texts are all there. The search isn’t working, so that’s next on the fix list. I’ll bet it’s something small. It usually is.

If you want to do this

You don’t have to build an app to get most of the benefit. Here’s what I’d tell a friend:

1. Export first. Run Google Takeout on Google Voice before you touch anything. Ten years of texts are worth an hour of downloading.
2. Practice on a new number. Get a cheap number at a carrier like Telnyx and live on it for a while before porting the one that matters.
3. Do the 10DLC paperwork early, and use your real documents. Make sure your business name and tax ID match your IRS letter exactly. Put SMS terms and a privacy policy on your website before you apply.
4. Check your two-factor codes. Some services won’t text codes to VoIP numbers. Move important accounts to an authenticator app before port day.
5. Push, don’t poll. If you build anything, have the server tell the app about new messages. Polling caused my worst bug.
6. Test against a fake carrier. A staging copy that fakes the carrier’s API caught more bugs than anything else, and it never sent a single real text.
7. Watch the whole path, not just the server. My worst failure produced no errors at all. A monitor that tests the outcome (does a VIP call actually ring?) would have caught it in minutes instead of days.

Was it worth it?

My business number is mine now. It’s not a feature of somebody else’s free product. Every text I’ve sent or received since September 15 sits in a database I own. My phone makes a different sound for each person. My AI assistants can help with my texts without scraping anything. And when a stranger calls while I’m under a desk, a robot politely takes a message for about seven cents.

I’d do it again. I’d just put the monitor in on day one.

← Back to the inbox

The scariest failure isn't the loud one. It's the one where everything looks fine.

Why let an AI answer?

When I’m under a desk at a client’s office with a screwdriver in my hand, I can’t take a call. Voicemail works, but most people hate leaving one. What I wanted:

Once the number lived at Telnyx, that became possible. Telnyx has a feature called Call Control: every time a call comes in, Telnyx asks your server what to do with it, and your server answers with commands like “answer,” “transfer” or “hang up.” My texting server already handled Telnyx’s texting webhooks. Now it handles calls too.

The rules

Here’s what happens when you call my business number today:

1. On my VIP list? Family and the people I most need to hear from (31 right now) ring straight through to my phones, showing their real caller ID, just like a normal call.
2. Blocked? The call ends. Spammers don’t get an AI to chat with.
3. Everyone else gets an AI receptionist. It says up front that it’s an AI and that the call is recorded, then takes the message: who you are, what you need, how to reach you.
4. I get a summary in Telegram right away, plus a digest every morning at 7:30.

At night (10 PM to 7 AM), only family and urgent VIPs ring through. Everyone else gets the AI. I can change any of it from Telegram with commands like /vip, /block and /night, with no app or website needed.

Want to see the whole route a call takes? The diagram on the section front page has a “Someone calls” button.

The three silent days

The move from “my office phone system answers my number” to “my server answers my number” happened on the night of September 21. The AI part worked. Calls got answered and messages showed up in Telegram.

What nobody noticed: the old path had been doing two jobs the new one didn’t.

• The VIP list stopped ringing through. My family and best clients were getting the AI too.
• The voicemail safety net was gone. If the AI couldn’t start, the call just ended. If my server was down, there was nothing behind it at all.

Nothing errored. Nothing alerted. The calls looked handled. It took three days before a full review of the phone setup on September 24 turned it up.

The fix

On September 25 I told Leo to fix all of it, and it went in that night:

• VIP ring-through is back, transferring the call to my office phone system so it rings my phones. That surfaced a sneaky gotcha. By default, a Telnyx transfer shows the number that was dialed as the caller ID, not the person calling. My phone system saw “my number calling my number,” didn’t recognize a VIP, and sent my VIPs to voicemail. The transfer now passes along the real caller’s number.
• If the AI can’t start, the call goes to my regular phone system and voicemail instead of hanging up.
• If my server is down, Telnyx falls back to a small backup responder on the office phone system, so a call always lands somewhere.
• Spending caps on outbound calls, so a bug can’t run up a bill.
• A monitor that checks the whole call path every five minutes and pings my Telegram the moment anything drifts from how it should be.

That round shipped with 100 of 100 tests passing. The monitor is the part I care about most. The three silent days weren’t caused by bad code. They were caused by a change with nothing watching it.

What it costs

In the last 30 days, the AI receptionist answered about 155 calls across all my lines, for a total of $10.55. That’s about seven cents a call to never get interrupted by an extended-warranty robot again.

Once it's your app, "I wish it could…" stops being a complaint and turns into a to-do item.

Version 0.3, the same night

Version 0.2 went live at 6:27 PM on September 23. Version 0.3 went live at 7:38 PM. That hour was mostly testing, because the features had been built that afternoon alongside everything else. Here’s what I got.

A sound for every person. Each contact can have its own notification sound and vibration. Under the hood, the app makes a separate Android notification channel for each person, since that’s the only way Android lets an app give different alerts different sounds. Now I know who’s texting before I look.

Priorities and VIPs. Every contact is Standard, Priority or Silent, and anyone can be a VIP. VIPs sit at the top of my People list and always count as priority.

Reminders that don’t quit. If a priority text sits unread, the app reminds me again, on a schedule I set per person (every 10 minutes, up to twice, for example). Reading it or replying cancels the reminders. They survive the phone rebooting, too.

Quiet hours and snooze, so all of the above leaves me alone at night unless it really matters.

A setup checklist. Android has a handful of settings that silently wreck alerts: battery optimization, exact alarms, Do Not Disturb access. The settings screen checks each one and gives you a Fix button that jumps straight to the right system page.

The upgrade trick

Here’s a small detail I love. The first rebuilt app had been installed under a test package name. Instead of switching to a “proper” name and making me log in again and redo every contact’s sound, the real release kept that package name and signing key. Android treated each new version as an update, so my login and all my per-contact sounds carried over every time. Not pretty, but it worked every time.

Goodbye to the old app

For about a day I had two apps on my phone with the same name: the old React Native one and the new one. Same name, different insides. I asked Leo to get rid of the old one, and it came off both phones. Its 7-second polling stopped for good on September 24.

The desk page

Most of my day is at a PC, so on September 24 the same server got a desktop page: texts and calls in one list, call and voicemail cards right inside each conversation, keyboard shortcuts for everything, and a Ctrl+K search box for jumping to anyone. It went through 31 automated checks before going live that night.

Voicemail moves in

My office phone system already turned voicemails into emails and transcripts. On September 26 the app got a Voicemail button with a badge. Tap it to play a message, read the transcript, call back or mark it heard. There’s a second tab for “AI took a message,” which you’ll understand after the next message.

The same weekend the app learned to open on my main line by default. My server sends from more than one number, and the app had been guessing.

The full test

On September 26 we test-called every line I own and sent test texts across all of them: 22 calls and 9 texts. It found one problem I never would have guessed. One of my phones had been sitting on manual Do Not Disturb since the 22nd, so calls rang silently and showed up as missed. No app bug. Just a setting. That’s exactly why the setup checklist exists.

The app was never the hard part. The hard part was everything the app assumed would just work.

Step one: stop asking, start listening

The old app asked the server “anything new?” every 10 seconds. The fix is to flip it around and have the server tell the app the moment something arrives.

On September 19 Leo shipped a web app version of the app built that way:

• Server-Sent Events keep one quiet connection open to my server, and a new text shows up on screen the instant it lands. No polling, so no redraws stomping on my typing.
• Web Push wakes the phone when the app is closed. Leo wrote that part from the spec and checked it against the official test vectors instead of trusting a library.
• A login per device that I can revoke one at a time, contact names, sharing photos into the app, and a Back button that actually goes back.

It passed 46 of 46 acceptance checks, and a real text from my main number to my 4040 line arrived in under six seconds.

“Never miss an alert”

Two days later I’d used it enough to know what was still wrong. I wanted different sounds for different people. I wanted reminders that keep nagging until I answer. The way I put it to Leo:

A web page can’t do that on Android. It can’t give each contact its own sound, and it can’t keep reminding you after you swipe it away. A real Android app can. So the design became a shell: a small native Android app that hosts the web app, plus the native pieces a web page can’t do:

• a background service that holds the live connection to my server,
• real Android notifications with reply right from the notification,
• per-contact sounds, reminders and quiet hours (that’s message 06).

September 23: one afternoon

On the 23rd I told Leo the app was still buggy. It still deleted what I was typing sometimes, it jumped around while scrolling, and Back exited the whole app. Leo stopped patching and rebuilt it.

At 2:34 PM it snapshotted the live code. Then three builder agents worked in parallel: one on the server, one on the web app, one on the Android shell, all working to one written contract for how the pieces talk.

A fake phone company for testing

You can’t test a texting app by spamming real texts all afternoon. So the rebuild came with a staging harness. It’s my real server code, wired to a fake Telnyx that never touches the network. Send a text and the fake carrier answers like the real API, then a second later plays back a “delivered” receipt through the same code path real receipts use. It can also inject fake incoming texts and fake picture messages on command.

Every screenshot of the app in this story comes from that harness, with made-up people in it.

What testing caught

The server review turned up 18 bugs in code that had been “working.” A few favorites:

• Vanishing texts. In a conversation over 500 messages long, the app loaded the oldest 500. New texts were in the database but never showed up.
• Lost-forever texts. The server marked an incoming webhook as handled before saving the message. If the save failed, Telnyx would retry, the server would say “already got it,” and the text was gone for good.
• “Retry” that didn’t. The retry button never actually re-sent anything.
• A freezing server. A few slow network calls blocked everything else while they waited.
• A sneaky picture. A picture message pretending to be a certain kind of image could have run code inside the app with one tap. Now only normal photo and video types open inline, and everything else downloads as a plain file.

Then putting it all together on real phones caught more. My favorite: the app and the server got into a “mark as read” argument that fired 647 times in two minutes. Another: Android never told the web page it was in the background, so the page kept marking texts read and cancelling my notifications while the phone sat in my pocket.

Shipped

By the end of the afternoon: 43 server tests, 72 web app tests, 28 of 28 checks on my Pixel and 27 of 28 automatic checks on my TCL (the last one, replying from the notification, I checked by hand). At 6:27 PM, version 0.2 went live on both phones, with a rollback copy saved in case it didn’t hold.

It held. So that night, we kept going.

The fastest way to learn what you want is to use something that isn't it.

The question

My first instinct on port day was to skip apps entirely. I live in Telegram, so I asked for a bot that would let me text through it “so I don’t have to deal with a weird app.” That works for alerts. It’s miserable for actual conversations with forty different people.

So at 3:29 that afternoon:

Orion came back with a short list of open-source options. Fifteen minutes later:

Forking an open-source app

Orion found VoIPSuite-Mobile, an open-source texting app from a privacy project. It’s built in React Native, and it already knew how to talk to a VoIP carrier. Orion forked it, pointed it at my own server, and named it KCCS Text.

Within two days it was on both of my phones. By September 17 it was at version 0.7.0, with a desktop-and-mobile “business hub” growing next to it on the server. For a forked weekend project, that’s fast.

Then I actually used it all day.

Bug one: the silent alarm

The point of a texting app is knowing when a text arrives. On September 17 we found that Android had been quietly marking every single notification from the app as SILENT. The notification channels had been created with sound and vibration off, and once Android creates a channel, the app can’t simply change it later.

The fix was new channels with new IDs, plus a lesson I’d keep relearning for the rest of this project: on Android, notification settings are sticky. If you get them wrong on the first install, you have to version your way out.

Bug two: the text eater

The one that drove me nuts showed up the next day:

Here’s what was going on. The app checked the server for new messages every 10 seconds, and each check redrew the screen, including the box I was typing in. Phone keyboards keep a little “work in progress” state while you type (the word being autocorrected, the swipe in flight). Redraw the box under it and the keyboard loses its place. Letters vanish. Sometimes a whole sentence.

Orion’s answer was a redesign, version 0.6.1. When Leo reviewed it, it caught a trap: the “new” version had the same internal version number as the 0.7.0 already installed, and it was missing picture messages and saved drafts. Installed over the top, it would have been a downgrade wearing a new coat of paint.

The audit

That night I told Leo what I really wanted. Roughly: make this the ultimate texting app for my number, not a half-baked thing I can kind of put up with.

Leo’s audit was blunt:

• The whole app was one giant file on a React Native version from 2022.
• There were three copies of the code drifting apart in different folders.
• The 10-second polling was the root of the typing bug, the battery drain and the slow delivery.
• Patching it would mean fighting an old framework forever.

The recommendation: freeze the old app, stop polling, and rebuild around a server that pushes new messages the instant they arrive. That’s message 05.

Moving a phone number is mostly forms. Then one morning, it's just done.

Can’t I just forward it?

The first question in September was the lazy one: could I leave the number in Google Voice and just forward everything to Telnyx? No. Telnyx won’t attach its texting service to a Google Voice number. To own the texting, I had to own the number. That meant porting.

The 10DLC boss fight, round two

This round, Orion (ChatGPT’s Codex) drove the paperwork with me. On September 8, in one long evening:

1. ID check. Telnyx verified my identity with a photo ID and selfie, which unblocked the account.
2. Business details. We caught that a draft of my business info didn’t match my actual IRS letter, threw it out, and used the real paperwork instead. Mismatched tax details are one of the most common reasons a 10DLC brand gets rejected.
3. Brand submitted at 4:51 PM. Verified at 8:47 PM.
4. Website pages. The carriers want to see SMS terms and a privacy policy on your site, including how STOP, START and HELP work. Those went up on my business site the same night.
5. Campaign submitted at 9:02 PM as “Customer Care,” which is exactly what my texting is: talking to people who already know me.

The fees, so you know what you’re in for: $4.50 for the brand, $15 for each campaign review, and $10 a month for the Customer Care campaign. The campaign sat in “pending” with the carriers for two days, then went active on September 10.

The letter

Porting a number runs on a Letter of Authorization (LOA). It’s a signed statement that you own the number and want it moved, plus proof, usually a screenshot or bill showing the number on your account. Google Voice also has you unlock the number for transfer first.

By September 11 the port order was filled out and waiting on one thing: my signature. We asked for the earliest activation slot: Tuesday, September 15, at 10 AM Mountain.

Port day

There’s no fanfare when a number ports. Nobody calls you. At some point in the morning, the old carrier lets go and the new one picks it up.

Here’s the message I sent my AI teammate at 10:42 that morning:

Ten years in Google Voice, gone from the app in one quiet refresh. Texts to my number were now landing on Telnyx and getting passed to my server. The first test texts went in and out. It worked.

The 2FA scare

One thing did rattle me that afternoon. A service I use refused to send a login code to my number, with an error saying it doesn’t send codes to VoIP numbers.

That’s a real thing to plan for. Some banks and apps won’t text verification codes to internet-based numbers. If you do this, check where your two-factor codes go before port day, and move the important ones to an authenticator app.

Six hours later

The number was mine. The texts were reaching my server. There was just one problem: I had nothing good to read them on. That afternoon I asked the question that started the second half of this story.

Rule one of moving something important: build the new house before you move out of the old one.

Don’t practice on the real number

Porting a phone number is a one-way door. Once my number left Google Voice, there was no clicking “undo” if the new setup couldn’t ring my phone. So the plan was to build everything on a new number first, live with it, and only move my real number once the new system had proven itself.

The night of June 27

Starting just before midnight on June 27, Leo (Claude, my main AI teammate) and I built a full business phone system from nothing:

• A carrier account at Telnyx. Telnyx sells phone numbers, calling and texting to businesses, and every piece of it has an API. A number costs about a dollar a month.
• A phone system on a rented cloud server. It runs FreePBX, a free, open-source office phone system built on Asterisk. That’s the software behind a lot of real business phone systems. It runs on a small cloud server that costs about $10 a month.
• A new business line, (719) 530-4040, wired through Telnyx into that phone system. Today an AI receptionist answers it.

By morning I had extensions, voicemail, call routing and a number that rang. It was the kind of thing a phone company used to charge a monthly fee per desk to set up.

That same night we also pulled my entire Google Voice history out with Google Takeout and turned it into a searchable archive with three copies. Whatever happened next, those ten years of texts were safe.

Making it trustworthy

A phone system that usually works is worse than none, because you stop checking. The next few sessions were all about the boring parts:

• Voicemail to email, so a message never sits unheard in a box I forgot about.
• Follow-me to my cell, with a “press 1 to accept” check. Without that, my cell’s own voicemail would grab the call and the phone system would think I’d answered.
• Telegram alerts for every missed call and new voicemail. Telegram is my pager. If it buzzes, something needs me.
• A fire drill. We called in, left messages and made sure every alert actually fired before calling it done.

First texts on my own server

On July 5, the first real milestone for texting: a text sent to the 4040 number landed on a little webhook service, which forwarded it to my Telegram and email. For the first time, a text to my business was data I had, not data Google had.

Receiving worked right away. Sending ran into a wall with a very boring name.

Meet 10DLC

In the US, if a business wants to send texts from a regular 10-digit phone number, the carriers want to know who you are first. The system is called 10DLC (10-digit long code). You register:

1. A brand: your business, verified against your tax records.
2. A campaign: what kind of texts you’ll send (customer care, marketing, reminders) with examples, opt-out wording, and a privacy policy on your website.

Skip it and your texts get filtered or blocked. My first try at registering the brand bounced with “Account unverified,” because the Telnyx account itself hadn’t passed an identity check yet. My account balance was sitting at $4.51 and slowly draining, and I wasn’t ready to fight that fight yet.

So I let it sit for the summer and had my AI teammates do the homework instead: what texting really costs, how 10DLC works, exactly how you get a number out of Google Voice, and what the alternatives were. Four research write-ups later, the answer was clear. The practice rig worked. The next step was the paperwork, then the real number.

Free is a great price until you want to do something the free thing doesn't allow.

A great deal, for a while

I got my business number through Google Voice in November 2015. It was perfect for a one-person shop. One number rang every device, texts showed up on my phone and my PC, voicemails got transcribed, and it cost nothing.

So everything went there. Customers. Vendors. Family. Every “your appointment is tomorrow” and “your package was delivered.” When I finally exported it all in June 2026, the archive held:

• 86,252 texts
• 10,900 calls
• 3,014 voicemails
• 5,045 contacts

That’s not a phone number anymore. That’s the memory of my business.

The wall

The trouble started when I wanted my computers to help with any of it.

I run a lot of automation. I have AI assistants that read my email, track my tasks, and draft replies. They can do all of that because email has real tools for it. Google Voice has none. There’s no official API for texts or calls. Your own software can’t read your own messages.

There was one crack in the wall: Google Voice can forward received texts to email. So for months I had a scraper pulling text notifications out of Gmail every day. It was held together with tape:

• Sent texts were invisible. Email forwarding only covers what comes in. Half of every conversation, my half, never made it out.
• Group texts arrived empty. The notification email for a group message has no message in it. Just “you have a new message.”
• Sending was off the table. Google’s rules don’t allow automated sending, and I wasn’t going to risk the number on a hack.

I had a note in my to-do list for a long time that said, more or less, find an unofficial way to scrape Google Voice until I migrate. “Until I migrate” was doing a lot of work in that sentence.

The part that kept me up

The bigger worry wasn’t convenience. It was this: the number my whole business depends on was tied to a free consumer Google account.

If that account ever got locked, or Google changed the product, or decided to retire it the way Google retires things, the failure would be silent. Clients would call, nobody would answer, and they’d never learn why. When I sat down and listed the single points of failure in my business, that number was near the top of the list.

I didn’t need Google Voice to be bad. I needed it to stop being the only door.

What I wanted instead

Written down, the wish list was short:

1. Own the number, with a carrier that sells to businesses and has an API.
2. Own every message, sent and received, in a database I can search and back up.
3. Let my tools in. My AI assistants should be able to read a text and draft a reply without anyone screen-scraping anything.
4. Never miss anything. Whatever replaced Google Voice had to be at least as reliable on my phone as Google Voice was.

Number 4 turned out to be the hard one. Next up: before I touched my real number at all, I built a whole phone system on a brand-new number to practice on.