Most questions here were actually asked — by users, in issues, or in conversations. Comparison answers carry dates, because other tools change.
Claude Code and Codex already read files in my repo. Why would I need this?
They do — and for a symbol-level coding task, raw access is often enough. StashBase is not trying to replace filenames, symbols, imports, or an agent’s native code search.
A codebase is also a knowledge base: architecture, decisions, documentation,
and product context live beside the code. Wiki Pages give people and agents a
reviewable map of that material, while searching by meaning helps retrieve ideas
expressed with different words. An AGENTS.md rule can force-read documents
you already know matter; it cannot discover which of 300 files is relevant to
today’s question. That discovery step is what the index adds. The
five-method comparison covers when
raw access is enough and when it isn’t.
For a narrow coding task, the agent’s native tools may be all you need. Use StashBase when you want the code, docs, and decisions to become a navigable, searchable context layer that can be reused across agents.
Doesn’t searching by meaning add latency?
No. Indexing runs incrementally in the background. Keyword search and normal file access remain available while it is being prepared; once it is ready, searching by meaning gives agents a faster way to discover relevant Sources.
What does the embedding cost?
Cents, on your own key. Search by meaning stays off until you add an OpenAI or OpenRouter key under Settings → Advanced → Search by Meaning; Default Agent credits do not cover it. OpenAI’s small embedding model costs $0.02 per million tokens (as of August 2026), so indexing even a large personal library costs cents. Keyword search and every local workflow need no key.
What about the “LLM wiki” / knowledge-graph idea?
It’s a great method — and it needs a workbench to live in. In StashBase you can have an agent organize and structure your documents into exactly that kind of wiki. The deliberate design choice is that the workbench isn’t tied to one representation: a knowledge graph, a Markdown wiki, or plain folders all sit on the same base — a general-purpose document library with search infrastructure underneath.
How is this different from NotebookLM?
NotebookLM is genuinely good at what it’s built for: a bounded set of sources you upload into a notebook, grounded answers with citations, and Audio Overviews — the generated-podcast feature, which StashBase has no equivalent of. For a self-contained project you’re happy to upload, it’s polished.
The difference is where the material lives. NotebookLM asks you to upload copies into per-notebook collections and maintain that second library by hand; when a file changes on disk, the uploaded copy doesn’t. That one-step upload sounds trivial — but maintaining a second copy of your library is the part people quietly stop doing. StashBase inverts it: your folders are the library. Files stay in place, the index is derived from them, and adding material is just saving a file where it already goes.
The other difference is the AI. NotebookLM is Gemini, inside Google’s app. StashBase brings your own agents — Claude, Codex, Antigravity, Cursor, any MCP client — to one shared library, so you don’t switch tools per scenario or re-upload the same material into each tool’s separate knowledge base. You can think of it as a local-first NotebookLM that’s agent-agnostic — the full comparison goes deeper, and AnythingLLM vs NotebookLM covers the most common self-hosted swap. (As of August 2026.)
Why not Obsidian?
Obsidian is an excellent personal knowledge base for humans — if your notes are Markdown and you live in it, keep living in it. The difference is that StashBase is agent-native: the index, the MCP server, and bounded file operations are the product, not plugins. Obsidian’s AI plugins exist and keep improving; as of August 2026, in our own use they still felt bolted on rather than built in. StashBase is also not Markdown-only — PDFs, DOCX files, and images join the same searchable library.
Does StashBase upload my files?
No. StashBase opens folders where they already are and does not upload, sync or copy them to a StashBase server. The text and indexes it prepares from them stay in app storage on your computer. What does leave is what an agent reads: when you chat, your prompts and the file content the agent reads go to that agent’s provider. What local-first means in StashBase lists each case.
Does StashBase work offline?
Partly. Browsing, editing and keyword search run on your machine and keep working without a connection. Agent chat and search by meaning need one, because the models run with the provider. The PDF and OCR components also need a download the first time you use them.
Does StashBase sync between devices?
No. There is no built-in sync. Projects are ordinary folders, so the sync or backup tool you already use can carry them, and each computer prepares its own index from the files.
Where does the Default Agent send my prompts?
Through StashBase’s hosted gateway and OpenRouter to the model you select, Anthropic’s Claude Haiku 5.5 by default. The gateway does not log or store prompts or replies. OpenRouter and the model provider process them under their own policies. If you would rather keep traffic under your own account, use Claude Code, Codex or Antigravity, which send traffic straight to Anthropic, OpenAI or Google under your own account. Details are on the privacy page.
Is StashBase open source?
Yes — Apache-2.0, the whole app: github.com/liliu-z/stashbase. It runs on macOS, Windows, and Linux.