← Open LocalMind

LocalMind — User guide

A private AI workspace that runs entirely inside your browser tab.

No server · no API keys required · nothing leaves your device
On this page

Getting started

Open LocalMind. The first time, it downloads a small default model (LFM2.5 230M, about 140 MB) and keeps it in your browser’s cache, so later starts skip the download. The welcome screen says which model is ready and offers a few example prompts. Tap one, or type a message and press Enter. The model, your chats and your memory stay on your device.

LocalMind welcome screen with the sidebar, the ready model and four example prompts
The welcome screen: the ready model, example prompts, and the sidebar with your recent chats.
Gemma 4 E2B answering a translation request, tagged On-device
A chat answered on the device (note the “On-device” tag).

Choosing a model

The model picker sits above the message box. It lists the models already on this device first, then any partly downloaded, then the rest with their download size, so you can see what will load straight away.

Settings → Models adds any Hugging Face ONNX repo or GGUF URL, or your own Ollama / LM Studio server. The full list is in FEATURES.md.

The model picker open, grouped into On this device and Download
The picker: what is on this device, then what needs a download.

The sidebar and ⌘K

The sidebar holds New chat, Chat, Image, Voice, Models and Library. More opens Compare, Batch, Diffuse, OCR, Vision, Clone and Folder. Search and Research stay as toggles above the message box, so you can add web results to any message.

The sidebar with More expanded
The sidebar, with More open.

Press ⌘K (CtrlK on Windows and Linux), or click Commands, for a palette of every action: type a few letters and press Enter.

The command palette filtered to the Settings commands
The command palette, filtered by “set”.

Compare — the model council

Open More → Compare, tick 2–3 models, type one question, and Send. LocalMind loads each model in turn (one at a time — WebGPU can’t hold several at once), collects every answer, then the last model loaded acts as judge and picks the best.

Model council: two answers and a judge verdict
Two models answer; the last card is the judge’s verdict.

Deep Research

Turn on Research above the message box and ask a question. LocalMind drafts several sub-queries, searches the web, reads the top sources, and writes a cited report with a sources list. It needs a web-search provider; if none is set, Research opens the setting for you.

Settings opened at the web search provider, with a note that Deep Research needs one
No provider yet: Research opens Settings at the search provider.

Batch

Open More → Batch and put one prompt per line (use Shift+Enter for new lines). Send runs them in sequence. Turn on Auto-inject {{previous}} to chain each answer into the next prompt.

Self-improving skills

The agent can save reusable skills — named recipes like “cite sources” or “be concise.” Once saved, a relevant skill is retrieved and applied automatically in later chats, so the assistant gets better the more you use it. Skills live in Library → Memory under the skill filter.

Library Memory tab filtered to saved skills
Saved skills (sample data), applied when relevant.

Library: memory and chats

LocalMind keeps a private, on-device memory (a local vector store). The agent can store facts; you can attach documents (PDF / DOCX / TXT) with the 📎 button or add a whole Folder (under More), and they become searchable context. Open Library → Memory to browse, filter, search or clear it.

Library Memory tab listing facts, preferences, skills and document passages
Everything the assistant remembers (sample data), filterable and under your control.

Library → Chats searches the full text of every saved chat, not only the titles.

Library Chats tab with the search monsoon matching one saved chat
One search across every chat (sample data).

Sync to a folder (permanent, cross-device)

By default all data lives in this browser. In Settings → Data, connect a folder and LocalMind mirrors everything — skills, memory, documents, conversations — to it as plain files. Point that folder at a git / iCloud / Dropbox directory and your data follows you across devices. API keys stay in the browser and are never written to the folder.

Settings, Data tab: data folder sync
Settings → Data: folder sync and backups.

Images and text diffusion

Image in the sidebar makes pictures on your GPU with a 4B diffusion model, or through your own stable-diffusion.cpp server. The picture shows large beside the controls, with a Download PNG button; earlier ones stay in the strip below.

Image screen: controls on the left, a watercolor of a paper lantern over a misty lake on the right
A real on-device generation: 768×512, four steps, seed 11.

More → Diffuse is an experimental way to write text: instead of word by word, the whole answer starts as masked tokens and “denoises” into place.

Diffuse mode mid-run, part of the answer still masked
Text diffusion, caught halfway: the masked parts are still in the fog.

Light, dark and phones

Settings → General → Appearance picks light, dark, or your system’s theme. On a phone, the ☰ button opens the sidebar, and the message box gets the full width with attach and voice stacked beside it.

LocalMind in the dark theme
The dark theme.
LocalMind on a phone: the welcome screen, with attach, memory and voice stacked beside the message box
On a phone.

Privacy

Models run on your device via WebGPU. Your conversations, memory, and documents never leave your machine. The only thing that touches the network is a web search — and only when you choose Search or Research, using a provider you configure. Once a model is cached, it loads from your browser with no download.