GemiHub Desktop

Chat supports one-click normal/wide sidebar layouts, compact manual Markdown exports, and automatic history retention. Manual notes use YYYYMMDD-HHmmss_Chat title.md; the default history limit is 100 and 0 is unlimited.

As AI gets smarter, the value of your data keeps growing.
GemiHub Desktop is the IDE for ideas—an open workspace that gathers thoughts, notes, reference materials, activity records, and even secret tokens on your own computer, then turns them into context for AI.

日本語 README

Get it from the Microsoft Store · Privacy Policy

Start with the work, not the prompt

Most AI tools start with a chat box. GemiHub starts with what you're reading. Open a document, select the passage that matters, and turn the surrounding work into grounded context for Chat, memos, or a Workflow. You do not have to leave the source and reconstruct its meaning in a prompt.

The best interface to an AI isn't syntax. It's intent. GemiHub treats actions such as “ask about this selection,” “memo this,” “organize these notes,” and “run this Workflow” as the interface. Files, selections, and the current Workspace carry the context; the user supplies the intent.

Why this app exists

GemiHub Desktop brings together three ideas developed in earlier projects:

• the fast document viewing, editing, and source-linked memo experience of mdwys;
• the RAG, OKF, AI Chat, Dashboard, and Workflow features of GemiHub;
• the LLM CLI and local-model integrations explored in obsidian-llm-hub.

The goal is a compact everyday Markdown application that can use AI deeply without making an online service mandatory. It is distributed as a single executable with no runtime dependency; the current Windows amd64 build is under 20 MB.

Files remain ordinary local files. Reading, editing, memos, Dashboards, Canvas, Base, Kanban, history, and Trash work without AI. When needed, the same Workspace can connect to OpenAI, Gemini, Vertex AI, Anthropic, Codex, Antigravity CLI, or a local LLM.

No single model is the product. Whichever provider is configured—a hosted API, a coding CLI already installed on the machine, or a model running locally—it reads the same Workspace and reaches the same tools. A configured CLI can power Chat, draft and revise Workflows, and run the LLM steps inside a Workflow, so changing providers is a settings change rather than a migration.

GemiHub provides a complete environment for using AI together with MCP servers, Agent Skills, OKF bundles, and Plugins.

What you can do

• Use it as your default Markdown app. Associate .md files with GemiHub, open them quickly from Explorer or Finder, and switch between Preview, WYSIWYG, and Raw editing.
• Find notes by meaning. Build a local RAG index over Markdown, text, PDF, and supported media instead of relying on filenames or remembering where a note was written.
• Let AI clean up the rough work. Ask Chat to rewrite, summarize, classify, or reorganize notes, then review file changes before applying them.
• Keep questions and impressions connected to their sources. Add source-linked memos while reading Markdown, PDF, EPUB, HTML, or technical books, and return from a memo to the original passage.
• Automate information gathering. For example, a Workflow can take an English technical article, translate it, and create a readable infographic note.
• Run daily work from one screen. Use Kanban and Calendar for tasks and schedules, record miscellaneous activity in Timeline, and quickly access frequently used files. Password-protect account details and retrieve them only when needed.
• Turn conversations into a daily record. Ask Chat or the Discord bot to “memo this” and it can save the answer or key points directly to the Workspace system Timeline.
• Give AI accurate product knowledge. Publish your own app documentation as an OKF bundle so Chat can answer quickly from a curated LLM wiki.
• Connect work systems. Use an MCP server to turn Git history into a daily report, or a Plugin to back up and synchronize files with services such as Google Drive.

A workspace that grows with plugins

GemiHub Desktop is not limited to built-in document and AI tools. Install plugins from Settings → Plugins to add complete, task-focused applications to the same local Workspace:

The same screen also supports portable Agent Plugins v1.0.0. Preview and install a public GitHub package to add its Agent Skills and stdio or Streamable HTTP MCP servers. Packages are pinned to a commit and can be disabled, updated, or removed independently from GemiHub UI plugins. Updates are previewed before installation, and an MCP server must be tested before it can be enabled; approval is cleared automatically if an update changes its connection details.

• Audio Score turns audio into playable sheet music with ML pitch detection, source separation, MIDI import and export, PDF export, and optional AI chord analysis.
• Accounting adds Beancount-compatible double-entry bookkeeping, transaction entry, validation, and financial reports while keeping the ledger in plain text.
• Ronginus runs structured, role-based debates between multiple configured AI or local models, then collects conclusions and votes.
• Google Drive Sync connects an entire Desktop Workspace to a GemiHub Google Drive root with previewable pull/push operations and per-file conflict resolution.

AI is optional. Without an API key, cloud account, or network connection, GemiHub remains a local document and knowledge workspace.

Voice input

In voice input settings, select the speech recognition API option to choose OpenAI, whisper.cpp, an OpenAI-compatible API, or Gemini 3.5 Transcribe. Gemini and OpenAI reuse the matching API key registered in AI settings, regardless of the active chat provider. When no matching key is configured, a speech-only key can be entered. Registering an AI key automatically takes priority. Transcribe supports two connections:

For lower-latency dictation, choose Live transcription with OpenAI, Gemini API, or Vertex AI. Interim text appears in the Chat draft while you speak. Speech settings also support question, line-break, exclamation and send commands, plus replacement rules such as daily note => /daily.

• AI Studio: use a Gemini API key.
• Vertex AI: reuse the current Google login and Project ID from AI settings, automatically following changes. No separate speech connection settings are needed. Enable the Vertex AI API and ensure access to the model. This uses gemini-3.5-transcribe-preview in global.

Set the language to auto, ja-JP, en-US, or another supported language code. Recording is limited to five minutes including retained clips. After recording stops, the transcript is inserted into the draft. The connection test sends a short silent recording; API usage may incur charges. Cloud Speech-to-Text V1 configuration is no longer needed.

Saying the send phrase with an empty draft ends listening instead of sending. A recording with no speech in it is never transcribed, and the notices about retained recordings or errors can be closed with their × button.

Answers can also be spoken: turn on Read answers aloud automatically in voice settings or in the tool menu next to the Chat input, where the speaking rate can be adjusted as well. Starting the microphone also starts voice mode, which reopens the microphone after each answer; the × beside the microphone stops only that automatic reopening.

References: Gemini API, Vertex AI.

Screenshots

Review AI changes before they touch your files

Chat can propose edits across Workspace files. Inspect the exact diff, choose individual hunks when needed, and apply only the changes you approve.

Ideas stay linked to their sources

Select text in Markdown, PDF, or EPUB and create an anchored memo alongside the document. Highlights navigate to memos, and memo quotes navigate back to the source.

A workspace shaped around the project

Arrange PDFs, HTML research notes, Markdown documents, and tools in rows or columns, then save dashboards as portable YAML files.

Knowledge remains discoverable

Browse every document with memos, ordered by recent activity.

Every edit has a way back

Review saved file versions and idle checkpoints, compare them with the current file, and restore an earlier state when needed.

Architecture

GemiHub Desktop is built with Go, Wails, Deno, Vite, React, Wysimark-lite, and pdf.js.

The desktop shell provides local filesystem access within a user-selected workspace. The React frontend renders documents and workspace tools, while AI providers, local CLIs, MCP servers, skills, plugins, and YAML workflows form an extensible intelligence layer.

Supported formats

• Documents: Markdown, plain text, HTML, PDF, EPUB, and images
• Workspace files: Dashboard, Base, Kanban, JSON Canvas, and Workflow YAML
• Encrypted files: self-contained .encrypted files that retain the original format

Safety model

• Workspace APIs and AI file tools are limited to the selected Workspace directory. Files opened from elsewhere are limited to their explicitly opened Files directory; traversal through .. or symlinks is rejected in both scopes.
• Up to 50 versions are saved before a file is overwritten, and deleted files can be restored from Trash.
• AI-proposed edits and renames require confirmation.
• Plugins declare permissions such as files, storage, network, and llm.
• File Widget text auto-saves after editing pauses. Encrypted text is re-encrypted with the session password and auto-saved; binary previews remain read-only.

Install

On Windows, install it from the Microsoft Store, which keeps the app updated automatically.

Everywhere else, download a binary from the GitHub Releases page. Deno and Go are not required at runtime.

Available release artifacts:

• gemihub-desktop-linux-amd64 (WebKitGTK 4.1)
• gemihub-desktop-linux-arm64 (WebKitGTK 4.1)
• gemihub-desktop-darwin-arm64
• gemihub-desktop-windows-amd64.exe
• gemihub-desktop-windows-arm64.exe
• gemihub-desktop-windows-amd64.msix (the Microsoft Store package, also installable directly)
• gemihub-desktop-windows-arm64.msix (the Microsoft Store package, also installable directly)

The Linux binaries use WebKitGTK 4.1 and support distributions such as Debian 12/13 and Ubuntu 24.04 or newer. Install the runtime library if it is not already present:

sudo apt install libwebkit2gtk-4.1-0

Each release also includes THIRD_PARTY_NOTICES.md. The same notices are available in the app under Settings → General → Third-party notices.

On Linux and macOS, make the downloaded file executable:

chmod +x gemihub-desktop-linux-amd64

The macOS binary is currently unsigned, so clear its quarantine attribute before first launch:

xattr -d com.apple.quarantine gemihub-desktop-darwin-arm64

Quick start

1. Launch GemiHub Desktop and choose a local workspace directory.
2. Click + Add Widget or drag files into the window.
3. Arrange sources in rows or columns.
4. Enable AI in Settings and configure a provider if you want AI features.
5. Add a file with @file, or select text in a document and choose Ask AI from the context menu, to give Chat grounded context.

Files can also be opened through the operating system's Open with action or as startup arguments:

gemihub-desktop note.md research.pdf book.epub

Development

Requirements:

• Deno 2.9 or newer
• Go 1.25 or newer
• Wails v3.0.0-beta.12 (go install github.com/wailsapp/wails/v3/cmd/wails3@v3.0.0-beta.12)
• Wails platform dependencies for your OS

Install dependencies and run the web UI:

deno install --allow-scripts
deno task dev

Run the desktop app in development mode:

deno task desktop

Type-check and build:

deno task check
deno task build
deno task desktop:build

Developer Tools are enabled in desktop builds. Press Ctrl+Shift+I (Cmd+Option+I on macOS) to open the WebView inspector.

Build Microsoft Store MSIX packages on Windows with the Windows SDK installed:

wails3 task windows:msix ARCH=amd64
wails3 task windows:msix ARCH=arm64

Before Partner Center submission, replace the Identity name and publisher in build/windows/msix/AppxManifest.xml with the exact values assigned to the reserved Store product.

Acknowledgments

GemiHub's built-in Markdown, Base, and Canvas Agent Skills include documentation adapted from kepano/obsidian-skills. The corresponding file-format support in GemiHub was independently implemented from the publicly described formats and behavior; it does not incorporate Obsidian source code. Canvas support follows the open JSON Canvas specification.

We are grateful to Steph Ango (@kepano), the project contributors, and the maintainers of JSON Canvas for making their work available to the community. See Third-Party Notices for copyright and license details.

GemiHub's WYSIWYG Markdown editor uses takeshy/wysimark-lite, a lightweight fork of portive/wysimark. We thank the Wysimark authors and contributors for the foundation they made available under the MIT License.

GemiHub Desktop is built with Wails (wailsapp/wails). We thank Lea Anthony and the Wails contributors for making the framework available under the MIT License.

GemiHub is an independent project and is not affiliated with or endorsed by Obsidian.

MCP permissions and read-only tools

MCP tool calls require approval by default. The dialog shows the server name, tool name and execution arguments. Choose Allow once, Always allow this tool, or Deny; dismissing the dialog denies execution. Server settings provide an Always approve option (off by default) and a removable allowed-tool list. Workflow command and mcp nodes can use confirm: "false" to skip MCP approval for that node, including automatic runs. Later MCP App interactions follow the server settings again. Read-only mode allows built-in reading, listing and searching, and blocks built-in writes (including timeline append). External MCP and skill/workflow operations retain their own permissions. Connection tests disable editing while running and display errors for correction.