A kitchen laptop running a local AI model is a shared brain. Anyone who sits in the same computer login can read old chats, open attached files, and reuse whatever model is already loaded. Local means the prompt may stay off a vendor’s servers. It does not mean private from the people who share your house.
Omar sat down at 9:40 p.m. The MacBook still had a crooked lemon sticker on the lid, and the chat app was still open from homework. Maya, 11, had asked for a science-fair paragraph on fermentation. The sidebar listed 11 threads. Third from the top was titled “week 6 notes.” He clicked it because it looked like a grocery list. The first assistant reply was a six-line summary of Sam’s therapy journal, the 2,800-word paste Sam had dropped in at lunch. Maya had not gone hunting. She used the same macOS user Omar never logged out of.
This page is about history, files, and who can reach the app’s network port. Menu names below are as of writing (August 2026).
A family Mac is not a team server
One kitchen laptop is a shared brain: chats, dropped files, and a listener on a local port. Sharing a login, sharing the machine with separate users, and running a real team host are three different setups. Write down which one you have before you invite a partner, a kid, or a coworker. You will also learn what 127.0.0.1 vs 0.0.0.0 means on cafe Wi-Fi, how to print a household rules card, and how to run a two-curl check for who can hit the port.
One laptop, one memory
A local model keeps the prompt off a vendor’s GPU if you stay in local mode (the hosted-vs-download guide’s cloud toggle still exists). The prompt still lives in RAM, on disk in whatever the app stores, and on the screen. Anyone in that OS user can scroll the sidebar, reopen “week 6 notes,” and read the six lines. They can drag the next PDF from Desktop. Local answers “did this leave the house.” It does not answer “who else in the house is this.”
Three buckets sit in that one window. Chat history: titles, full turns, attached files. Model files: GGUF-class blobs under ~/.ollama/models on macOS, or the Windows path Ollama documents under your user folder. The network listener: as of writing, Ollama binds 127.0.0.1 port 11434 unless you change OLLAMA_HOST. LM Studio’s developer server is a cousin, often on 1234, with a “Serve on Local Network” switch that binds off localhost. llama.cpp’s server can do the same. None of those local HTTP APIs ask for a household password. Ollama’s auth docs say no authentication is required on http://localhost:11434. That helps you. It is a hole if the bind is wider than this machine.
Maya used the lemon-sticker login because it was already open and homework was due. Sam used the only chat on the only user, named the thread like a journal heading, and went back to stirring. Nobody needed a villain. The shared brain was enough.
Rule of thumb: If two people can sit in the same OS user, they share the sidebar. Name the user the way you would name a filing cabinet.
Share a login, share a machine, or run a real host
People say “we share Ollama” and mean three different setups. Write down which one you actually have before you invite a partner, a kid, or a coworker.
| Mode | What you share | Who sees the chats | Who can hit the API |
|---|---|---|---|
| Share the login | One OS user, one runner app, one ~/.ollama | Anyone at that keyboard, including a child on homework | This machine only, until someone changes the bind |
| Share the machine | Separate OS users, same laptop, same fan | Each user’s own chats if they signed out | Still localhost per session, unless a service was rebound for everyone |
| Real team host | A box with an owner, backups, and a written update plan | What that product logs, on purpose | Named people on a private network, with auth in front of the port |
Share the login is what Omar had. Fine for a solo laptop. Wrong the moment a second person types. The cheap fix is social: New Chat, delete the private thread, title the rest with a name. The better fix is an OS user per adult. macOS Fast User Switching is slower than “here, use mine.” It is also how journals stay in Sam’s home folder and homework stays in Maya’s.
Share the machine still shares heat, RAM, and disk. the hardware guide already told you a 16 GB laptop gets unhappy with an 8B-class model and Chrome. If Maya loads a model at 7:50 p.m., Omar’s evening compile waits. Agree who may run a model the same way you agree who may render video.
A real team host is a different animal. Someone owns updates (the update guide). Someone owns backups of the model directory. Someone can say who may call the API. A kitchen Mac with OLLAMA_HOST=0.0.0.0 so three laptops can curl it is a toy server. Toys are allowed on a Saturday. They are not an admin plan on Monday when an intern pastes a customer CSV.
Who can hit this port
127.0.0.1 is this computer talking to itself. Apps on the same Mac can reach Ollama. A phone on the Wi-Fi cannot. 0.0.0.0 means listen on every network interface this machine has. The cafe DHCP address is an interface. The home LAN address is an interface. Ollama’s FAQ documents how to set OLLAMA_HOST to 0.0.0.0:11434 when you want other devices on a network to connect. The docs are a warning label with an on-switch, not a dare.
As of writing, that local API still does not ask the caller to log in. A second device that can open http://192.168.x.x:11434/api/tags can list models. Generate and chat endpoints are the same HTTP you already use. LM Studio’s network-serve page is blunt: any bind other than 127.0.0.1 exposes the server beyond localhost, and they recommend turning authentication on. If you need a phone in the house to talk to the Mac, that is a LAN-night decision with a firewall rule, not a default you leave on when you close the lid and walk to a coffee shop.
Cafe Wi-Fi is the sharp case. You sit down, the laptop joins Airport_Free, and a leftover launchctl setenv OLLAMA_HOST 0.0.0.0:11434 from last weekend’s demo is still in the environment. The runner comes up with the lid. Anyone else on that SSID who scans common ports can find 11434. They do not get your Apple password. They get an unauthenticated model endpoint, which will run whatever prompt they send against whatever weights you loaded, using your RAM and your battery. Home is calmer and not empty: guest phones, a kid’s friend with a laptop, a neighbor on an open guest network. You need to know whether the listener is loopback or LAN. Two curls, plus a look at who is on the SSID.
A household rules card
Print this, tape it under the trackpad or on the inside of the lemon lid. Change the names. Keep the bind line. The comments are part of the card so the next adult can read why the rule exists.
# KITCHEN OLLAMA CARD (print, tape under the trackpad)
# House: Omar / Sam / Maya. Stack: Ollama. As of August 2026.
# 1. Adults have separate macOS users. Maya uses "Kids", not Omar.
# 2. Chat titles start with a name: Omar-work, Sam-private, Kids-homework.
# 3. Before you hand the laptop, New Chat. Delete a private thread.
# 4. Journals, tax PDFs, and work decks stay out of a house window.
# 5. Bind stays 127.0.0.1:11434. Nobody sets OLLAMA_HOST=0.0.0.0
# without a Saturday talk the rest of the house hears.
# 6. Cafe, hotel, airport, school guest Wi-Fi: Quit Ollama in the menu bar.
# 7. If work wants "our home box," the answer is no. They buy a host.
# 8. Models live in ~/.ollama/models. Do not "free space" there on a school night.
# Who can hit this port? Run on the Mac that hosts the runner.
curl -sS -o /dev/null -w "localhost:%{http_code}\n" --max-time 2 http://127.0.0.1:11434/api/tags
# 200 means the app is up on loopback. Now try this laptop's LAN IP
# from a phone on the same SSID (Wi-Fi details in System Settings):
# curl -sS -o /dev/null -w "lan:%{http_code}\n" --max-time 2 http://192.168.1.47:11434/api/tags
# If the phone gets 200, you bound too wide. Clear OLLAMA_HOST, restart.
# macOS leftover from a demo (only if you set this on purpose):
# launchctl getenv OLLAMA_HOST
# If that prints 0.0.0.0:11434 and you did not mean LAN night, unset it and reopen Ollama.What that card is for: a 30-second read at the table. Line 2 is how Omar would have seen “Sam-private” instead of “week 6 notes.” Line 3 is the handoff. Line 5 and the curls are the only technical part. Flags and menu labels move. If curl fails, the app may use another port; check the runner’s status screen that week.
A small team is not the kitchen box
Three people in a shop will try this: one quiet desktop, a 12 GB model that “almost feels like ChatGPT,” OLLAMA_HOST=0.0.0.0, and a Slack note with the LAN IP. It works until someone needs an audit trail, until two prompts queue and the fan becomes the office soundtrack, until an intern tunnels the port with ngrok “for a client demo,” or until the update guide’s update lands a new default model and Tuesday’s summaries change voice. Ollama’s FAQ even shows ngrok and Cloudflare Tunnel examples. Those tools put localhost on the public internet. That is a product decision with auth and TLS, not a dinner-table experiment.
If the work is real, pick a host you can name. A hosted open-model API (Groq the inference company, spelling with a q, not xAI’s Grok; or Together, or Fireworks) gives you a key, a bill, and a retention page you can actually read. A small VM you rent, with a reverse proxy and a password, is the self-host version of that. The kitchen Mac can still be the place you try prompts that must not leave the house. It should not be where accounting, support, and a contractor all point their laptops. Write four lines on a sticky before you share a port at work: who owns the box, who may pull models, where the disk lives, what happens when the person with the lemon sticker is on vacation. If you cannot fill those lines, you do not have a team host. You have Omar’s laptop with extra DNS.
Worked example: 9:40 p.m., 11 threads
Here is the evening as a table you can reuse, with the numbers from Omar’s lid.
| Clock | What happened | What was shared | Fix that would have held |
|---|---|---|---|
| 12:40 p.m. | Sam pasted 2,800 words, titled the thread “week 6 notes” | The full journal, in the only OS user | Sam-private on Sam’s login, then sign out |
| 7:50 p.m. | Maya opened Ollama still running under Omar | 11 titles in the sidebar, including week 6 | Kids login, or Omar quits the app after lunch |
| 8:10 p.m. | Maya asked for a fermentation paragraph; model used the loaded 8B | Fan, RAM, and the same window Sam used | Fine, if the sidebar had no private titles |
| 9:40 p.m. | Omar clicked week 6, read a six-line summary | Therapy content, now in Omar’s head too | Delete thread after use; name chats; separate users |
The 10-minute fix that night: Sam created a macOS user. Omar deleted “week 6 notes” from his sidebar (the paste was already in Sam’s original file). Maya got the Kids login and a chat titled Kids-homework. Omar ran the localhost curl (200) and the LAN-IP curl from his phone (connection refused). Nobody set 0.0.0.0. They put the card under the trackpad. If a child is in the mix, read the kids-and-shared-devices guide for closed-chat age gates, Family Link, and Screen Time. That post is ChatGPT-on-an-iPad. This one is Ollama-on-a-laptop. The shared-brain problem is the same shape. The control is the OS user plus the port, because your runner may not ship a “family plan.”
Sharing a folder is not a policy
- Leaving one OS user open “because it is the house computer,” then being surprised the sidebar is a group chat.
- Titling private threads like grocery lists (
week 6 notes,stuff,Chat). - Setting
OLLAMA_HOST=0.0.0.0:11434for a living-room demo and never unsetting it. - Binding wide on cafe, hotel, airport, or school guest Wi-Fi, or leaving the runner up in a bag on those networks.
- Calling a LAN IP in Slack “the team server” with no auth, no owner, and no backup of
~/.ollama/models. - Tunneling 11434 with ngrok or Cloudflare because a FAQ snippet showed the command.
- Letting kids use a work login, or letting work use the kitchen box. the kids-and-shared-devices guide and the privacy paste test still apply; local does not cancel them.
- Deleting the models folder to “clean the disk” the night before a science fair (go back to the update guide).
One rule card on the machine
Sit at the actual table. Open the runner from the desktop-runners guide and the one-stack guide. Count the threads. Rename any that a kid or a partner should not open. Delete the one you would not read aloud. If two adults use the machine, make the second OS user now. Run the two curls. If the phone can see port 11434, put the bind back to localhost and restart. Quit the app once on a network you do not own, so the habit exists. Write the four team lines only if someone at work asked.
This series stops here on purpose. Hosted chat was the hosted-chat guide. Desktop runners were the desktop-runners guide. One stack was the one-stack guide. Offline jobs were the offline-jobs guide. Hardware was the hardware guide. Updates were the update guide. Sharing the box is the sharing guide. For model cards, licenses, and when a $20 closed chat is easier, switch to Open-source AI explained. For kids and closed products, the kids-and-shared-devices guide. More paths: Learn and the chooser.
Who can run it, what they cannot paste
- One OS user is one filing cabinet. Split logins or wipe threads before the laptop changes hands.
- Keep Ollama (and friends) on
127.0.0.1. Treat0.0.0.0as a LAN-night switch, never a cafe default. - Confirm with curls: localhost should answer, a phone on the SSID should not.
- A kitchen Mac with an open port is a toy. Work that needs sharing buys a host with a name and a bill.
- Tape the rules card to the lid. Title chats with people, not moods.
Series notes
This is the last part of Run open models from scratch (series code OS14). Previous: updating models without breaking your setup. Related: Open-source AI explained, privacy paste test, kids on shared devices, and Learn.
Sources
Product pages and docs used for bind defaults, auth, and network serving. Re-check the week you change a setting.
- Ollama FAQ (default bind
127.0.0.1:11434,OLLAMA_HOST, exposing on a network, ngrok and Cloudflare examples, model storage paths) - Ollama: Authentication (no auth required on local
localhost:11434) - Ollama (the app Omar’s house actually ran)
- LM Studio and LM Studio local server (developer server and the “Serve on Local Network” bind; they recommend auth off localhost)
- llama.cpp (engine under many desktop runners; its server can bind wide too)
- Groq (hosted inference for a real team seat; not xAI Grok)
- Together AI and Fireworks AI (hosted open-model APIs when the kitchen box is the wrong shape)
- AMS P29: Kids, family, and shared devices (closed-chat household layer)
- AMS P4: Privacy, run it yourself and AMS hosted vs download (location vs open weights)
- Analytics Made Simple: Learn
Keep going
Same lessons in your feed
Short diagrams, hooks, and weekly tutorials on Substack, Instagram, X, and Facebook.
