Use Ollama, LM Studio, llama.cpp or Jan from Self Chat

Self Chat can talk to a model running on your own computer. You run Ollama, LM Studio, llama.cpp's server or Jan, and the page sends each message from your browser tab straight to that program. The live page puts it this way: "Those messages never touch the internet at all."

These steps match the live site as of October 14, 2026. The feature shipped on October 2, the same day Self Chat got its name and its own page. Self Chat saves your chats only where you choose covers how Self Chat stores chats and handles cloud models.

Step 1: Connect storage

Self Chat keeps each chat only in a place you pick, and the chat opens once you connect one. Two choices work on the live site today:

• Google Drive. "This site can see only the files it made in your Drive."
• A folder on this computer. "This site sees only that folder; nothing is uploaded." This one needs Chrome, Edge or another Chromium browser on a computer.

The page lists OneDrive, Dropbox and Box under "Coming soon".

Pick the folder if you want the whole chat to stay on your machine. A local model keeps your messages off the internet, and the saved copy still goes wherever you chose. A chat in Google Drive is a file Google holds like any other file there.

Step 2: Start a model server

Any of the four works. Self Chat looks for each one on its usual port on localhost:

All four offer an OpenAI-compatible chat API, and the page asks each for its model list at /v1/models.

For a first model, try the one behind the default pick on tryingopen.com: Liquid AI's LFM2.5-VL 3B, which we run on palm-sized Jetson boards. Our default model runs on a palm-sized Jetson Orin Nano covers that setup. Self Chat's model picker leaves out the default pick, so running your own copy is how you use LFM2.5-VL 3B in Self Chat. On Hugging Face, the Use this model menu on Liquid's GGUF page gives this Ollama command for the 4-bit Q4_K_M build, a 1.67 GB file:

ollama run hf.co/LiquidAI/LFM2.5-VL-3B-GGUF:Q4_K_M

Liquid's model card lists a 32,768-token context. It recommends the model for single-turn, low-latency tasks and advises against long-context, reasoning-heavy ones.

Step 3: Let the site talk to your server

Each server decides which websites may call it, through a browser rule called CORS. The dialog in Self Chat shows the steps below with the address of the page you're on filled in. On the live site that address is https://www.tryingopen.com.

Ollama

Ollama answers only pages on your own computer until OLLAMA_ORIGINS includes the site. On a Mac, run this, then quit Ollama from the menu bar and open it again:

launchctl setenv OLLAMA_ORIGINS "https://www.tryingopen.com"

On Windows, run this, then quit and reopen Ollama:

setx OLLAMA_ORIGINS "https://www.tryingopen.com"

On Linux, start it with the variable set:

OLLAMA_ORIGINS="https://www.tryingopen.com" ollama serve

LM Studio

Load a model, open the Developer tab and start the server, then turn on "Enable CORS" in its settings.

Jan

Go to Settings > Local API Server, turn on CORS and start the server.

llama.cpp's server

llama.cpp's server allows any site by default, so starting it is enough. The dialog's example loads a small Gemma 3 build:

llama-server -hf ggml-org/gemma-3-4b-it-GGUF

Step 4: Find your model

In Self Chat's sidebar, press Use a model on this computer. A dialog titled "A model on your computer" opens. Press Find it and the page checks the four ports.

For each server that answers, the dialog lists its models. It skips embedding models. It marks models that read pictures with "sees images", a detail it gets from Ollama's and LM Studio's own model information. Press Use beside one and the chat switches to it. Your local models also sit at the top of the model picker, in a group called "On your computer".

If a server is running but refuses the site, the dialog says it "won't talk to this site yet" and shows that server's steps from Step 3. If nothing answers, it lists all four servers with their steps. A server on a different port goes in the "Running it on another port?" field.

Step 5: Chat

Each reply from your model carries an "On your computer" mark. Its tooltip reads: "Answered by a model on your own computer, called from this page: no server saw it". Self Chat saves it to your Drive or folder like any other chat.

If a reasoning model writes its thinking inside <think> tags, the page moves that text into the reasoning panel.

Browser caveats

• Chrome asks once for "Local network access", its permission for a site to reach devices on your network. Allow it.
• Safari usually refuses an https page's call to http://localhost. Use Chrome, Edge or Firefox instead. The dialog says this when it spots Safari.
• Firefox reaches local models. The folder option needs a Chromium browser, so pair Firefox with Google Drive.

For a model server, the page's security policy allows only localhost and 127.0.0.1. A typed address pointing anywhere else gets "Only a server on this computer works here (localhost or 127.0.0.1)." That keeps a chat from going to another machine because of a mistyped or pasted address.

What the browser remembers

Your browser keeps only the server's address and its model names, so the models are back in the picker on your next visit. The dialog's Forget button (it names your server, as in "Forget Ollama") clears them.

What "never touch the internet" covers

The privacy policy says these messages "never reach our servers or the internet." The saved chat sits only in the storage you picked in Step 1.

Web search is the one exception. Self Chat keeps search off until you switch it on in the composer. With search on and a local model picked, your latest message goes to tryingopen's server for one lookup, and, in the privacy policy's words, "only the results go back to your computer." The results go into your model's prompt and show on the reply as sources. The lookup goes through OpenRouter to a search provider, and that search sits outside the zero-retention requirement. How web search works on tryingopen explains that path.