Chat
Ask anything. Answers are composed on this device from your documents, your memory and real local tools — every claim traceable, every limit stated. Load an optional model to unlock free-form writing.
Learn
Eight short lessons on getting real work out of a browser AI — what it can do, what it must not pretend to do, and how to make it answer well. Everything is interactive: each lesson drives the real page.
Your progress
0 of 8Lessons
Which reasoning approach?
Auto picks for you. When you want to steer, this is the short version.
| Job | Pick |
|---|
Glossary
Tasks
Give a goal with several parts. Lantern writes a plan, retrieves evidence, computes with real tools on this device, then checks itself — and tells you which parts it could not settle.
Paste raw output from any model. Tool calls are read from fenced tool blocks or JSON with a tool key, validated against the schema, then executed here on your device.
Documents
Add the text you want answers grounded in — notes, contracts, specs, revision material. Files are read in this tab and never uploaded. Answers that use them cite them, with the passage attached.
Add a document
Test retrieval
0 passagesCosine similarity over 64-dimension hashed embeddings — the same vectors retrieval uses. Useful for spotting duplicated material.
Paste a draft answer. Each claim is checked against your documents and labelled supported, partial or unsupported. It flags risk — it cannot prove a claim true.
Memory
Facts you want carried between conversations: preferences, deadlines, context. Each one is scored for relevance, recency and importance, and stored on this device. Everything here is yours to edit or erase.
Teach it something
0 suggestionsRecall
0 memoriesAll memories
What it learned
Lantern adapts from your ratings, your corrections and the facts you save. Nothing is inferred behind your back: everything below is visible, editable, and erasable in one click.
How answers are shaped
defaultsCorrections you have given
0When you correct an answer, the correction is stored and shown again the next time a similar question comes up.
Ratings
Privacy and storage
Reasoning lab
Put one problem through any of the approaches in the literature and read the trace, the branches and the verification passes. With an on-device model loaded these drive real generation; without one they run the same methods over your documents and tools, and label the output honestly.
Approach
autoResult
composedTrace
Prompt studio
Write prompts that survive contact with a model: worked examples, injection defence, meta-prompts, and reusable reasoning templates. Everything runs locally on the text you paste.
Heuristic scanner for known attack shapes: instruction override, role injection, delimiter escape, prompt exfiltration, encoded payloads, egress URLs. It lowers risk; it does not make a model safe on its own.
Buffer-of-thoughts style templates: saved reasoning structures to hand to a model. Stored locally, editable, exportable.
My tools
Build a reusable tool from the parts of this system — an instruction, a reasoning approach, the capabilities it may use and the shape of its output. Saved tools appear in Chat. They export as JSON.
New tool
creatingYour tools
Developer instruments
The numbers behind an AI feature: context budgets, cost and latency, tool-declaration validity and Model Context Protocol messages. All computed here, no calls out.
Paste a tool declaration (OpenAI-style functions or MCP tool input schemas) and check it before a model ever sees it.
Builds and validates JSON-RPC 2.0 messages for the Model Context Protocol, then runs them against a local mock server so you can see the exact response shape. No network involved.
Request
Response
—Model & privacy
Lantern works with no model at all. Loading one is optional, happens only when you press the button, and runs the weights on this device through WebGPU.
This device
checking…On-device model
not loadedHonest limits
- No model, no invention. Without a model Lantern composes answers from your documents, your memory and computed tool results. It does not write free-form prose and it says what it could not do.
- Small local models are weak. A 0.4–2 GB on-device model is far below a hosted frontier model. The working is always shown so you can judge the answer rather than trust it.
- Retrieval is lexical plus hashed vectors. BM25 and 64-dimension hashed embeddings are fast and private, but they are not a trained semantic encoder.
- Not advice. Medical, legal and financial output is general information. Check anything that matters with a professional.
- Nothing leaves this device. No account, no cookies and no analytics on this page; the only optional request is the model download above.