StashBase is evolving toward three connected capabilities:
A VS Code-like Document Workbench, a Codex-like Agent Panel, and a local RAG layer for document retrieval.
StashBase should provide a workbench for browsing, reading, editing, navigating, and organizing documents across ordinary local folders. Like a code workbench, it uses a file tree, persistent tabs, quick navigation, and format-appropriate surfaces, but it remains centered on documents rather than code. It works with the user's existing files without replacing them with a database, block editor, or proprietary storage model.
The built-in Agent Panel works against an explicit library or folder scope. Before a document is opened, Chat is the primary working surface; once a document appears, the same Chat adapts into a side panel alongside the source. It is a convenient client of StashBase context, not a separate AI workspace and not a replacement for external Agent clients.
Opened folders become retrievable context. The local RAG layer prepares difficult formats, supports exact and meaning-based retrieval, and delivers source-grounded evidence to Agents. It should explain readiness and failures clearly without becoming a search or vector-database administration console.
Code repositories give Agents strong lexical structure through paths, symbols, imports, and stable identifiers, so iterative grep and file reads can often locate relevant code without a persistent semantic index. Document libraries are less predictable: a question may not share the wording of its sources, and evidence may span long-form files, OCR, or transcripts. StashBase therefore treats preparation, a persistent meaning-based index, and source-grounded retrieval as one first-class RAG layer instead of relying on exact terms alone.
AI Index needs a source of embedding capacity, and StashBase strongly steers every user to set one up at first run — an unindexed library has a degraded Agent — while stopping short of forcing it. Two sources are available: a hosted StashBase account with free monthly usage as the low-friction default for most people, and a bring-your-own OpenAI/OpenRouter key for advanced users.
Recommend, don't lock. Signing in should unlock StashBase's hosted service, not unlock computation the user's own machine can already do — so browsing, editing, preview, and keyword search must never be gated behind a remote login. The setup dialog leads hard toward enabling indexing and has no casual dismiss, but it offers a deliberate, low-emphasis exit ("Skip AI Index for now") to a "basic mode". No-index mode is a real, supported state, not a peer presented with equal weight — so the exit is a plain, low-key link, a window-local "for now" rather than a permanent opt-out. A fresh blank window makes the offer before a folder is selected. Skipping there covers the first folder opened so the same launch does not immediately ask twice; after that, the choice is remembered per folder in that window. Returning to a skipped folder stays quiet, while another folder or a fresh window re-offers indexing. The surviving local abilities are not advertised as a competing feature; the default guides everyone to enable.
Activation must not turn local files into something that needs the cloud to open. The governing rule: first use should choose an indexing source, but daily use must never depend on online auth to reach local files. Activation persists locally, while a skip follows the window-local context rule above; the app opens and serves its existing index offline; a network or service error never forces re-authentication; and when hosted free usage is exhausted, new AI Index updates pause while the existing index and Agent retrieval keep working. In basic mode the Agent still connects but flags, on first use or a failed retrieval, that indexing is off. Signing out or removing the key returns the library to the unactivated state.
The current direction favours contributions that improve:
- Markdown authoring and preview fidelity.
- The clarity and reliability of preparation, indexing, and retrieval.
- The usability and safety of the Agent Panel.
- The Document Workbench's everyday reading and maintenance workflows.
- Cross-platform reliability and an approachable contributor experience.
These themes guide prioritisation; they are not release commitments. Area-level work and its status live in the design documents.