Local Agent Studio

Local-first Agent Mode for Windows, Linux, and macOS.

A ChatGPT/Claude-style desktop app that connects your local and API-powered tools: Ollama, ComfyUI, web search providers, workspace files, local databases, and sandbox commands.

Local Agent Studio is built for people who want a normal chatbot interface, but with tools they control locally or through their own API keys.

Highlights

• Chatbot UI inspired by ChatGPT, Claude, and Open WebUI.
• Editable user messages with real context rewind to that point in the conversation.
• Ollama chat with streaming responses.
• Reasoning panel for models that return thinking traces.
• Reasoning flags with --think, --think high, and --no-think.
• Reasoning picker near the model selector for quick ChatGPT-style control.
• Multimodal attachments for images, audio, video, text, and files.
• Drag-and-drop and clipboard paste for images/files.
• Native image input for Ollama vision-capable models.
• Local PC date/time context so the LLM knows today's date without guessing or searching.
• Web search through SearXNG, SerpAPI, or Ollama Web Search.
• Search summarization instead of raw result dumps.
• Up to 3 web searches per user request.
• ComfyUI image generation with bundled workflows.
• Image model picker for Z-Image-Turbo, Flux.2 klein 9b, and Ideogram v4.
• Custom ComfyUI image models through user-defined workflow JSON files.
• Ideogram v4 effort selection: turbo, default, quality.
• Workspace file tools: create, read, edit, delete, and preview text files.
• File preview and download from the workspace panel.
• Local database creation from JSON/CSV-like objects: JSON DB, CSV export, and SQLite.
• Sandbox commands through subprocess or Docker mode.
• MCP server support for external tools.
• Runpod provider settings for remote Ollama-compatible and ComfyUI workloads.
• Tool permission controls for files, search, images, terminal, database, and MCP.
• First-launch workspace setup.
• Chat export/import as JSON.
• Manual update checks through a lightweight remote version.json file.
• Agent task queue for sending the next prompt while the current answer is running.
• Tool activity timeline for visible agent/tool steps.
• Image gallery with generated image history, full-screen preview, prompt details, and download controls.
• Artifacts for reusable generated documents, code, JSON, Markdown, and design drafts.
• Personal memory between chats, with on/off controls.
• Remote provider presets for local, Runpod, and LAN-style setups.
• LM Studio and llama.cpp provider presets through OpenAI-compatible endpoints.
• Project templates for Work, Dev, Fun, and Research workspaces.
• Plugin marketplace with official plugin cards, GitHub manifest installs, and MCP-backed extensions.
• Model capability badges for text, image, audio, and MCP support.
• Themes: system, light, dark.
• Languages: English, Russian, Ukrainian, German, Polish.

Screenshots

What It Can Do

Official Plugins

See OFFICIAL_PLUGINS.md for the GitHub Git Operator setup guide.

Ask it like a normal assistant:

What is the current price of Meta Quest 3S in Ukraine?

If web access is enabled, the agent can run search queries, collect context, and synthesize a readable answer.

Create files in the workspace:

Create a test text file with "letter 123" inside.

Generate images through ComfyUI:

Generate 3 product shots of a local AI desktop app, use Ideogram quality mode.

Create a database from objects:

Create a database named products from objects:
[
  {"name": "Quest 3S", "price": 299},
  {"name": "Quest 3", "price": 499}
]

Run a sandbox command:

/run Get-ChildItem

Download

Release builds are generated under release/.

Windows installer:

release/Local Agent Studio-0.2.1-x64.exe

Linux packages:

release/Local Agent Studio-0.2.1-x64.tar.gz

AppImage and .deb builds should be produced on Linux or CI with Linux packaging tools.

macOS packages must be produced on macOS or macOS CI:

release/Local Agent Studio-0.2.1-x64.dmg
release/Local Agent Studio-0.2.1-arm64.dmg

If you are publishing this project on GitHub, attach the installer to a GitHub Release.

Quick Start From Source

Requirements:

• Windows 10+, Linux, or macOS
• Node.js 20+
• npm
• Ollama for local LLMs
• ComfyUI for image/video workflows
• Docker, optional, for SearXNG and sandbox isolation

Install dependencies:

npm.cmd install

Run in development mode:

npm.cmd run dev

Build the frontend:

npm.cmd run build

Run the packaged-style Electron app:

npm.cmd start

Build the Windows installer:

npm.cmd run package:win

Build Linux packages:

npm.cmd run package:linux

Build Linux AppImage/deb installers on Linux or CI:

npm run package:linux:installers

Build macOS packages on macOS or macOS CI:

npm run package:mac

Local Services

Ollama

Start Ollama:

ollama serve

Pull a model:

ollama pull llama3.1:8b

You can keep the model setting as auto, or set a model manually in Settings. The app also includes quick presets for Gemma 4-style model names:

gemma4:e2b
gemma4:e4b

ComfyUI

Default endpoint:

http://localhost:8188

Bundled workflow files:

electron/workflows/image_z_image_turbo.json
electron/workflows/image_flux2_text_to_image_9b.json
electron/workflows/ideogram_v4.json

Supported image presets:

• Z-Image-Turbo
• Flux.2 klein 9b
• Ideogram v4

For Ideogram v4, the app exposes:

• turbo
• default
• quality

SearXNG

The repo includes a Docker Compose file for local search:

docker compose -f docker-compose.searxng.yml up -d

Default endpoint:

http://localhost:8080

Search Providers

Local Agent Studio supports:

• SearXNG for local/self-hosted search.
• SerpAPI through your own API key.
• Ollama Web Search through your own API key.

Environment variables:

$env:SEARCH_PROVIDER="searxng"
$env:SERPAPI_API_KEY="..."
$env:OLLAMA_API_KEY="..."
$env:SEARXNG_BASE_URL="http://localhost:8080"

You can also configure providers inside the app Settings.

Configuration

Settings are stored locally in the Electron user data folder.

Useful environment variables:

$env:LOCAL_AGENT_WORKSPACE="C:\Users\You\LocalAgentStudio\workspace"
$env:LOCAL_AGENT_THEME="system"
$env:LOCAL_AGENT_LANGUAGE="en"
$env:OLLAMA_BASE_URL="http://localhost:11434"
$env:OLLAMA_MODEL="auto"
$env:OLLAMA_THINKING="off"
$env:COMFYUI_BASE_URL="http://localhost:8188"
$env:SEARCH_PROVIDER="searxng"

Workspace

The Workspace tab lets you:

• Browse workspace files.
• Create new files.
• Read text files.
• Edit and save text files.
• Delete files or folders.
• Open the workspace folder in Windows.

The agent can also perform workspace actions from chat, for example:

Create notes/todo.md with a checklist for publishing the project.

Databases

The agent can create simple local databases from structured objects.

Outputs:

• databases/<name>.db.json
• databases/<name>.csv
• databases/<name>.sqlite, when the runtime supports node:sqlite

Example:

Create database named launch-list from objects:
[
  {"site": "Hacker News", "status": "planned"},
  {"site": "Product Hunt", "status": "later"}
]

Chat Commands

/search <query>

Force web search and answer synthesis.

/run <command>

Run a command in the selected sandbox mode.

Natural language image generation also works:

Generate a GitHub banner for this project using Flux.

Explicit image command examples:

/image local AI desktop app hero banner --count 3
/image product screenshot in a laptop mockup --model ideogram --effort quality

Project Structure

.
├─ electron/
│  ├─ backend/
│  │  ├─ attachments.cjs
│  │  ├─ comfy.cjs
│  │  ├─ config.cjs
│  │  ├─ data.cjs
│  │  ├─ files.cjs
│  │  ├─ llm.cjs
│  │  ├─ providers.cjs
│  │  ├─ sandbox.cjs
│  │  └─ search.cjs
│  ├─ workflows/
│  ├─ main.cjs
│  └─ preload.cjs
├─ src/
│  ├─ components/
│  ├─ App.tsx
│  ├─ i18n.ts
│  └─ styles.css
├─ assets/
│  ├─ github-hero-dark.png
│  ├─ github-social-light.png
│  └─ rendered-screen.png
├─ docker-compose.searxng.yml
└─ package.json

Development Checks

Build:

npm.cmd run build

Capture an Electron screenshot:

npx.cmd electron .\scripts\capture-electron.cjs

Interactive capture:

$env:LOCAL_AGENT_CAPTURE_INTERACT="1"
npx.cmd electron .\scripts\capture-electron.cjs
Remove-Item Env:\LOCAL_AGENT_CAPTURE_INTERACT

The screenshot is saved to:

assets/rendered-screen.png

Roadmap

• Finished ComfyUI job watcher that pulls generated images back into chat.
• Search source cards with extracted price tables.
• Model capability registry for text, image, audio, video, reasoning, and context size.
• Conversation history saved as local projects.
• SQLite/CSV/JSON table viewer.
• Ask-before-write approval mode for risky actions.
• Git tools: status, diff, commit summary, changelog.
• Workspace RAG/indexing.
• Plugin system for custom tools.
• First-run setup wizard.
• Signed icon and branded installer.

Security Notes

• subprocess mode runs commands on the host machine.
• Use Docker mode for untrusted commands.
• Workspace file operations are scoped to the configured workspace directory.
• API keys are stored locally in app settings or read from environment variables.
• Review tool output before relying on generated files, commands, or search results.

Contributing

Issues and pull requests are welcome. Useful contributions:

• New ComfyUI workflows.
• Better model presets.
• Search provider integrations.
• UI polish.
• Windows installer improvements.
• Documentation and demo videos.

License

Apache License 2.0. See LICENSE.