Changelog - VaultRAG

1.3.0 - Git, a phone remote, scheduled jobs, and more agents

VaultRAG grows from a chat with agents into a workspace that also holds your repositories, runs while you are away, and follows you to your phone. Highlights:

A full Git view. Type git or click the branch chip and the conversation's project folder opens in an editor-style Git view: a timeline of commits, a file tree with status badges, side-by-side or unified diffs, stage and commit, branches, stashes, worktrees and tags, compare any two sides, a three-way merge editor, and fetch, pull and push explained in plain words. It runs on the git you already have installed. Pop it out into its own window per repository.
GitHub built in. Sign in with your browser, borrow your gh login, or paste a token, and the Repositories panel lists everything you can reach, with push and read-only badges, invitations you can accept, one-step clone, publish a local folder to GitHub, remote branches that appear without a manual fetch, and pull requests you can list, create and check out. Tokens live in your operating system's keychain.
Git in chat. /diff shows the uncommitted changes, /review asks the model to review them (or a base branch, or a commit), and after every agent turn the reply notes what the agent changed in the repository.
Remote control from your phone. Turn it on from /remote-control, pair a phone with a one-time code, and a full web app runs over your own Tailscale network: chat with any backend, pick models and working folders, search and download vault documents, answer agent approval cards, review git changes, and manage scheduled jobs. Nothing goes through a relay or a third party; the server listens on localhost and only Tailscale Serve reaches it.
Scheduled jobs. /cron followed by a request in plain words ("every weekday at 5, remind me to review open PRs") has the model build a job on VaultRAG's scheduler. Jobs run a prompt or a script on the built-in chat or an agent at set times, land each run in its own conversation, choose what to do about approvals while unattended, and show up in a Scheduled Jobs pane on the desktop and the phone.
Any Agent Client Protocol agent. Alongside Claude Code, Codex and OpenCode, VaultRAG can now drive any CLI that speaks the Agent Client Protocol, with presets for Goose and Hermes Agent and a place to add custom commands. Their commands, models and modes appear in the picker, their sessions resume, and their permission requests use the same approval card.
A clearer setup wizard. The providers page lists OpenRouter and every coding agent with an install and login check, a toggle per agent, and a default for new chats. A new document search page puts the embedding choice in one place, with chunked hybrid search preselected and a single toggle deciding whether local AI or OpenRouter embeds your documents.
Deep research and goals on every backend. /research runs VaultRAG's planned, cited research sweep in Codex and OpenCode conversations and forwards to Claude Code's own research skill; /goal keeps the built-in chat working until a condition holds. Agent chats gained /clear, /effort and Codex /fast.
Help, one page per feature. help is now an overview plus help git, help remote-control, help cron, help agents and a dozen more, shared by the desktop and the phone.
Verify and Repair that understand chunks. The database report is graded into files, folders, summaries and embeddings, and the chunk index, separating problems Repair fixes from work the indexer will finish. Repair no longer discards the chunk index, and the background sweep cleans orphan rows and a drifted full-text table on its own.
Chat and analytics polish. Forked conversations keep their agent and branch the Claude session, HEIC photos attach directly, equation numbers render, the picker opens instantly even with hundreds of models, the input bar has a new icon set, the vault list view's detail column is configurable, and Analytics adds a cost timeline, a per-backend donut, OpenCode usage, and roll-ups by account.
Hardening. Safer quit and startup (in-flight replies are saved, a startup failure shows a window instead of vanishing), a stable OpenGL renderer by default on macOS with Metal opt-in, and a long list of watcher, chunk index, remote boundary and cancellation fixes from the release QA sweep.

1.2.0 - Claude Code, Codex and OpenCode agents, plugins, and hybrid search

The biggest release yet. VaultRAG now connects directly to the Claude Code, Codex and OpenCode coding agents, gives the built-in chat a working folder with real file tools, adds deep research and a plugin system for custom AI tools, and introduces chunk-based hybrid search. Highlights:

Claude Code and Codex in chat. Pick an agent from the same picker you use for models and your chat turns run through the real CLI on your machine, with live streaming, tool cards, and colored file diffs rendered in the conversation. You log in through your own terminal with the Claude or ChatGPT account you already have; VaultRAG never sees your credentials.
OpenCode as a third agent. The open-source OpenCode agent joins Claude Code and Codex in the picker, driven through its own local server, with its thinking variants, slash commands and provider logins, or the OpenRouter key you already configured.
Deep research. /research <question> plans several searches, sweeps the vault (and the web when the globe is on) over two rounds, reviews the findings for gaps, and writes a report with numbered citations and a Sources section.
A working folder and file tools for the built-in chat. Every conversation gets a private workspace, or a project folder you choose, where the model can read, write, edit, list, glob and grep files, fetch web pages, and, with an opt-in, run sandboxed commands, each step shown as a tool card with diffs and previews.
Agents can search your vault. An optional read-only MCP server, bound to your machine only and protected by a per-session token, gives agents semantic search, passage search, document content, outlines, and folder structure.
Supervised approvals. Agent actions outside the safe allowlist pause on an in-chat approval card: Allow, Allow for this conversation, or Deny. An explicit full-access mode is available when you want no prompts.
Per-conversation project folders. Each conversation works in its own directory, a jailed sandbox by default or any folder you choose, and restores its project, backend, and session when you reopen it.
Slash commands. Type / to autocomplete built-in and custom Claude Code and Codex commands, including a dedicated /usage panel with per-limit progress bars.
Agent model and effort controls. Codex model lists are fetched live from your account, Claude Code offers its current lineup with six thinking-effort levels and an optional 1M-token context window, and the chat globe governs agent web access.
Agent usage analytics. A new Analytics section computes usage and estimated cost from the CLIs' own local logs, with per-backend breakdowns, daily cost charts, and cache savings, covering all your local agent use.
Plugin system. Drop-in plugins add new tools the AI can call in chat, loaded in isolation with read-only vault access, a Settings tab, and an SDK with a sample plugin and a dev guide.
Chunk-based hybrid search. Index documents as chunks with adaptive sizing and per-document evidence scoring, so answers can be found deep inside long files, with a live chunking preview in Settings.
Chat and file manager polish. Image attachments for agents, sticky per-conversation backends, ZIP compress and extract, Open in Terminal, verified transfers to external drives, and new themes bringing the collection to 74.

1.1.5 - Windows on ARM, memory improvements, and safer file transfers

This release brings VaultRAG's local AI to Windows-on-ARM devices, makes long chat sessions dramatically lighter on memory, and hardens file moves to external drives. Highlights:

Native Windows on ARM. A dedicated arm64 build brings full local AI support to Snapdragon and other Windows-on-ARM devices. Local models now run natively instead of failing under x64 emulation, and updates automatically deliver the right architecture.
Conversation memory comes back. Clearing, switching, or closing a large conversation now releases its memory instead of holding it for the life of the app, and image caches are freed along with it.
Safer moves to external drives. Cut/paste and drag-and-drop moves verify every file landed intact at the destination before the original is deleted, writes are flushed to the device, and a cancelled or failed copy can no longer leave a partial file behind that looks complete.
Smooth long chats. Long conversations are virtualized and load in chunks, so scrolling, clearing, and switching stay responsive even with hundreds of messages.
PDF cover thumbnails. PDFs show sharp first-page previews in the grid and the resizable details panel, backed by a memory-capped cache. A new Settings toggle turns thumbnails off entirely.
Chat text size and navigation. Pick a chat text size in Settings or use Cmd/Ctrl +/-, jump between your questions with Shift+arrows, and enjoy steadier scroll anchoring during streaming and resizing.
Charts and math in chat. Models can emit plot blocks that render as real charts (copyable as images), and LaTeX rendering handles more constructs with correct baselines and spacing.
Local embedding models. Run GGUF embedding models locally for document search, alongside a new colorful chat text option and an always-on-top window setting.

1.1.4 - File templates, Golden Gate theme, and vault fixes

This patch release adds quick file creation, a new theme, and a round of vault and chat reliability fixes. Highlights:

New File templates. Create files directly in the vault from format templates instead of switching to another app first.
Golden Gate theme. A new warm theme joins the collection.
Vault indexing fixes. Copied folder contents are auto-indexed, deleting ignored folders cleans up their database entries, and hidden files can be revealed reliably.
Chat selection and rendering. Text selection spans message blocks correctly, drags copy without disturbing the conversation, and currency values and tables parse more reliably.
Windows runtime packaging. The installer always bundles the newest available VC++ runtime, fixing local AI startup on clean machines.

1.1.3 - Local model options and reliability fixes

This patch release expands the local model catalog and improves how VaultRAG displays file metadata, model limits, currency ranges, and Windows file actions. Highlights:

Vinci Piccolo local model. Vinci Piccolo 1.0 Q4_K_M is now available in the managed local model catalog for chat and document summaries.
Clearer local model limits. Context-window sizes and output limits are now reported separately, avoiding misleading maximum output values for long-context models.
Reliable currency rendering. Prices and ranges such as $499-$699, $8K-$15K, and $1.5M-$3M remain readable text in chat paragraphs, lists, and tables instead of being mistaken for math.
Native Windows Open With. Windows now uses the system Open With chooser, including for files that already have a default app.

1.1.2 - Chat tabs, pop-out windows, and local AI polish

This release adds multi-conversation chat with tabs and pop-out windows, better conversation history organization, and a round of local AI and chat rendering fixes. Highlights:

Chat tabs. Run multiple conversations at once, each with its own model and live streaming state. Open a new tab with Cmd+T / Ctrl+T; the tab strip stays hidden until you have more than one conversation.
Pop-out chat windows. Pop any conversation out into its own window for side-by-side use, styled to match the main window. Closing a pop-out re-docks it as a tab.
Better conversation history. Pin important conversations, search your history, and browse it grouped by date. Deleting a conversation that is open now asks for confirmation and closes its tab safely.
Smarter local AI errors and stats. Request timeouts now come with guidance to raise the timeout instead of a generic connection error, and chat stats report real token usage from local models instead of estimates.
Vision-aware attachments. Image attachments are now blocked up front for local models that can't see images, instead of being silently dropped.
Smoother streaming. Answers stream in live again when web search and other tools are enabled, without stray text flashing before a tool call starts.

1.1.1 - PDF image parsing fix

This patch release fixes PDF parsing for documents that include images, improving indexing and chat context for image-heavy PDFs.

1.1.0 - Local AI and more

The biggest update since launch. VaultRAG can now run entirely on your own computer, chat responses stream live, and the file manager, themes, and onboarding all got a lot more polished. Highlights:

Local AI. Run chat, summaries, and embeddings entirely on your own machine: no API key and nothing leaves your computer. Install models from a built-in catalog that flags which ones fit your hardware, with GPU acceleration (including CUDA on Windows) and a one-click setup that picks sensible defaults for you.
Live streaming responses. Answers now stream in as they are generated, for both cloud and local models.
Web search. Use OpenRouter's built-in web search or your own Tavily key to bring live results into chat.
Choose what gets embedded. Index documents using an AI summary for the best search quality, or embed document text directly for faster indexing, especially with local models.
Conversation history and search. Browse past conversations and search within a chat, with matches highlighted inline.
Richer chat. Math and LaTeX rendering, syntax-highlighted code blocks, inline images, and improved vision support for images, PDFs, and email attachments.
More themes and appearance options. Dozens of new light and dark themes, a font picker, a compact minimal mode, and an option to center the window when it opens.
File manager improvements. Progress indicators for large copy, move, and delete operations, Finder-style icons, a customizable vault toolbar, file previews, and a native macOS share sheet.
Smoother onboarding. A guided setup that lets you choose cloud, local, or a mix, with one-click local AI setup built in.

1.0.0 - Initial release

The first public build of VaultRAG. Highlights of this release:

Local vault. Point VaultRAG at any folder on your machine and it builds a searchable index of your documents in place.
Embedding-based search. Find documents by what they actually contain, with optional filtering by file or folder name.
Automatic summaries. Every document is summarized on import. Summaries can be regenerated, edited by hand, or cleared at any time.
Chat with your documents. Ask questions against your vault and get answers grounded in the documents you have indexed.
Bring your own model. Connect to any provider through OpenRouter using your own API key.
Customizable system prompt. Tune the model's tone, focus, and behavior to match how you work.
One-time license. Pay once for the app, activate on up to two devices, and pay your AI provider directly for tokens.

Have feedback or want to report an issue? Reach out at [email protected].