CursivCursiv local AI: capabilities and limits.
Cursiv is a free, open-source local AI application with tamper-evident notebooks, offline document retrieval and portable Windows workspaces. Use it for everyday questions and document research, or keep detailed observations with an inspectable correction history.
Windows 10/11, 64-bit. Apple Silicon Mac build is early. Initial downloads require internet access.
What is local AI?
A local language model generates answers on your computer instead of sending each prompt to an online model provider. Cursiv uses Ollama as its local runtime; it is an application around existing models, not a new foundational model. Its default is Qwen3 4B Instruct on supported setups.
Cursiv’s application code is MIT-licensed. Downloaded models have separate licenses and hardware requirements. “Free,” “local” and “open source” describe different things: check the app, runtime and model you actually use. Read the local versus offline guide.
Who is it for?
Researchers preserving observations, technicians recording readings, investigators organizing source material, independent workers keeping job notes, and people who want AI help alongside detailed records.
You can use local chat without keeping a field notebook. The notebook becomes useful when the history of a reading or correction matters. See the field research workflow for a specialized example.
AI for documents and notes on your computer
Import supported documents into Research Files, retrieve relevant text locally and ask questions with page references where available. Keep selected useful facts in memory and look up saved notebook records by supported dates, sites, samples and measurements.
Cursiv does not automatically reorganize your entire computer. File actions operate inside a folder you authorize. The imported library is currently shared across workspaces; project names are not isolated access boundaries. Memory is selective, not a perfect transcript. Retrieval can miss relevant passages, and a model can misread a supplied source.
Workspace file permissions · Current research workflow and limits
Tamper-evident notebooks: keep the original
Notebook entries are linked with SHA-256 fingerprints. A correction is a new linked entry rather than a quiet replacement of the old measurement. Checking the chain can reveal changes relative to a previously trusted record or export. Exports can also be signed and checked against a known device key.
Tamper-evident does not mean tamper-proof. A chain alone cannot establish who observed something or whether it was true; someone controlling the files may rewrite a whole chain. Preserve trusted fingerprints or signed copies and follow the documented verification procedure.
Calculations performed in code
Supported unit conversions, distances, dilutions and statistical calculations use deterministic tools. This separates those results from arithmetic improvised in model prose.
Unsupported questions still reach a model, and inputs or assumptions can be wrong. A successful tool calculation does not certify the rest of an answer. See measured checks and failures.
What works without an internet connection?
After downloading the runtime, compatible models and needed documents, local chat, local document retrieval, notebook records, selective memory and supported calculations can work offline. Cursiv’s offline mode blocks its connections to outside services. Model downloads, update checks, web search and optional cloud providers require internet access.
Choosing a remote provider sends the content used by that request to that provider. “Local AI” is not a blanket privacy promise for every option. Read the reproducible offline demonstration and privacy notes.
Portable AI on a USB drive or external SSD
A prepared Windows kit can carry Cursiv, its Ollama runtime, selected model files and chosen workspace data. A complete compatible kit can run from the drive; copying only an installer or notebook is not the same thing. The host still needs suitable Windows hardware, memory and drivers.
Portable data remains readable by design, and Windows or other software may leave host traces. Close the app and eject safely. What travels, what remains on the host and what was tested.
Start with a small local setup
- Check your computerWindows 10/11, 64-bit; 8 GB RAM and roughly 6 GB disk space are starting points, not promises for every model. Check hardware and model fit.
- Download and verify CursivUse the current release and its owner-signed manifest. Windows publisher signing and Apple notarization are still pending.
- Install a model, then try offline useSetup installs Ollama and downloads your model. Start with a question or import a document, then enable offline mode for local work.
Choose by your actual needs
Cursiv is newer software, and AI answers can be confidently wrong. There is no Linux installer; the Mac build is early and portable drive operation is Windows-only. Broader report comparisons, general conversational exports, comprehensive date-scoped aggregates and project isolation remain unfinished.
No first-time-user study or universal accuracy score is claimed. Compare AnythingLLM, LM Studio, Jan and Open WebUI, then read the technical evidence and limitations.
Source and scope
Product reference for the U106 application, reviewed October 26, 2026. Published by the Cursiv project. These are documented capabilities, not an independent review or a claim of universal model accuracy.
U106 verification · Model licenses · Notebook reference · Short overview