One caveat first. The hands-on observations below are one writer's experience, not a benchmark. The product details come from SurfSense's own documentation, which is developer-published material.
What SurfSense is, and where it runs
SurfSense describes itself as a free, open-source desktop app for the documents you already have. The Windows download is a 64-bit installer called SurfSense-Setup.exe, with Apple Silicon and Linux x64 builds alongside it.
The install documentation adds some fine print for Windows users:
- It requires 64-bit (x64) Windows.
- Windows on Arm runs the x64 build under emulation and hasn't been tested.
- A PowerShell one-liner is offered as an alternative to the downloaded installer. According to the docs, it checks files against the release's published checksums, and the installer must carry a valid SurfSense signature.
- Those are developer-described safeguards, not an independent audit.
- Local models need enough memory, and a graphics card helps. The docs say Windows machines with a graphics card use it through Vulkan, and NVIDIA, AMD and Intel cards all work. Integrated graphics count as the processor.
- The Apache 2.0 license means you can inspect, modify and contribute to the code.
The project's own repository states the license plainly as Apache-2.0. It also notes that the older hosted web app is being retired in favor of the desktop app. Anyone still using the cloud version should check the export deadline stated on the project's sunset page.
The NotebookLM half: Sources, chat and Studio
Faisal describes an interface that will look familiar to NotebookLM users: a Sources section, a chat panel and a Studio panel. Studio lists twelve output types: Markdown, Word, Slides, Spreadsheet, Web page, PDF, Mind map, Flashcards, Quiz, Podcast, Image and Infographic.
The project's repository says several of these are editable files rather than pictures. Slides come out as a .pptx, reports as a .docx and spreadsheets as an .xlsx. Podcasts and images need extra models. The repository says podcast voices come from an offline voice model, and image and infographic outputs need an image model plus a generation model. The project's quick start says you don't need Docker, a terminal, a GPU or a compose file. Video overviews, according to the repository, aren't built yet.
Her test results
- Grounding was imperfect. Before uploading anything, she asked "What is XDA?" SurfSense said its knowledge base had nothing on XDA, then gave a generic answer anyway. After uploading a document, an unrelated question about Miami weather got a general-knowledge answer. The app only noted at the end that her document didn't cover the topic.
- Summaries worked well. The document summary was fast, neatly bulleted and cited. Clicking a citation opened the source and highlighted the relevant passage.
- Studio was mixed. The mind map, a web page generated from her document, and slides all worked well. Flashcards, quizzes and podcasts repeatedly failed to generate. She gave no error messages or configuration details, so the cause is unknown.
The practical lesson is that a citation-heavy interface doesn't guarantee answers stay inside your documents. If you rely on strict source-only answers, as you might with legal, research or compliance material, check the answers yourself. A listed Studio format is also not proof it will work in your setup.
Models: choice is the headline feature
SurfSense's documentation lays out three ways to supply a model:
- On this computer. SurfSense downloads and runs a model itself. Once it's downloaded, chat can work with the network off.
- A server. Any OpenAI-compatible service or local server, such as LM Studio or Ollama.
- A ChatGPT or Claude plan. You sign in with an account you already have, with no API key, and usage counts against that plan's limits.
You pick a model per job. Only text generation is required. Image generation, image editing, audio and video are optional and can be skipped. The embedding model finds relevant passages in your documents, and the quick start warns it can't be changed after setup. Choose it carefully, and pick a multilingual one if your documents aren't in English.
Faisal names Qwen3 and Gemma 3 as local options, and ChatGPT Plus or Pro and Claude Pro, Max or Team as plan options. The official docs I reviewed confirm the categories but not those exact model names or tiers, so treat them as version-dependent. She also used a text model she calls "GPT-5.6 Luna", a label I couldn't verify against SurfSense's documentation.
The privacy claim, read carefully
The project's marketing calls SurfSense "air-gapped". The documentation draws a more careful line:
- Local indexing. Documents are parsed, chunked and embedded on your machine. The repository says the index lives in SQLite under ~/.surfsense, and that no account is needed.
- Local models. Nothing leaves the computer once the model is downloaded, unless a chat uses plugins.
- Remote models. Your questions and excerpts of your documents go to that server or provider.
- Plugins. A plugin's server receives what the agent sends its tools. The docs say it can't open your files or Sources directly, but the agent may include text from the sources ticked for a chat, such as a search query.
- Downloads. The first local model download asks permission to reach huggingface.co. SurfSense says it sends only the model name and your IP address, as any download would.
So "nothing leaves your PC" is true only for a fully local setup with no plugins. Faisal makes the same point about cloud models. The app does ask before contacting plugin servers, and you can read each plugin's publisher, authentication, connections, tools and privacy policy before connecting.
The Perplexity half: Agentic mode and plugins
Plugins work only in Agentic chats. In Basic (Q&A) mode the Plugins menu is greyed out. Each plugin is an MCP server that offers tools such as "search Google". There are three kinds: SurfSense's own, third-party, and custom MCP servers added by URL.
Agentic mode also has model requirements. According to the docs, the model must support tool use and have a context window of at least 32,768 tokens. Otherwise new chats stay Basic.
Faisal asked Agentic mode to research an event she'd attended. She says it ran eight Google searches in parallel, prioritized official sources, and returned a structured rundown of five announcements, their availability and community reaction. She judged it decent based on her own knowledge of the event. She doesn't name the event, so that result can't be independently checked.
Useful controls from the plugin docs:
- Each tool can be set to always allowed, needs approval, or blocked.
- Each plugin call shows as a step you can open to see the whole result.
- "Save to Sources" keeps a plugin result as a note so later chats can cite it.
The catch: plugin pricing
The app is free, but SurfSense's own plugins aren't. Faisal reports a 14-day trial, then $60 for the first year and $120 afterward. SurfSense's pricing page agrees with those figures. It adds that this is a single payment rather than an auto-renewing subscription, and the trial needs only an email address.
When a license lapses, only SurfSense's own plugins stop. The rest of the app keeps working, including chat, Studio, your index and third-party or custom plugins.
There is a discrepancy in the plugin list. Faisal lists ten official plugins, including Reddit and Instagram. The pricing page I reviewed lists Google Search, YouTube, TikTok, Google Maps, Amazon, Walmart, Indeed and a web crawler. Treat the exact catalogue as changeable.
Setup: how it goes
Faisal says setup was far easier than Open Notebook, which she says took hours. That's her personal comparison, not a measured one. The documented first run goes like this:
- Click Start setting up.
- Answer two opt-in privacy questions: check for updates, and share usage statistics. Nothing is sent until you choose.
- Pick a text model: local, server, ChatGPT plan or Claude plan. This step can't be skipped.
- Optionally pick image, editing, audio and video models, or skip them.
- Choose an embedding model, then click Finish.
- Add files or folders, or use "Add from the web…" or a new note, and start asking questions.
She says usage limits are generous. Local models are limited by your hardware, and subscription connections by your plan. She reports not hitting a limit on her ChatGPT plan.
Who should try it
SurfSense makes sense if you want NotebookLM-like workflows without uploading documents to Google, or if you want to choose your own model. It also suits anyone who wants editable Office-format outputs.
Skip it, or test carefully, if any of these apply:
- You need strict source-only answers. Faisal's tests showed leakage into general knowledge.
- You need reliable flashcards, quizzes or podcasts. Those failed in her trial.
- You want web research without paying. The Google Search and crawler plugins require a license after the trial.
- Your hardware is modest. Local model speed depends on memory and graphics.
The project's own comparison concedes that NotebookLM wins on audio quality and offers video, which SurfSense doesn't yet. That's the developer being unusually candid, and it's a fair summary. SurfSense isn't a drop-in clone. It is a private, flexible alternative that's still rough in places, and its best feature is that you decide where your data goes.
References
- I tried replacing NotebookLM and Perplexity with one open-source research app, and it does more than just answer questions XDA · 2026-11-20T17:30:22+00:00
- GitHub - JoeMakuta/SurfSense: Open source alternative to NotebookLM for teams. Join our Discord: https://discord.gg/ejRNvftDp9 · GitHub github.com
- Local models | SurfSense Docs surfsense.com