Build local AI agents, connect visual multi-agent workflows, and automate useful work from one approachable Windows desktop app.
Local Agent Studio is an open-source, local-first AI agent builder and visual multi-agent workflow automation app for Windows. It is designed for everyone—from people creating their first AI agent to developers prototyping private agentic workflows. Create agents, mix local and cloud models, connect tools, watch every workflow step, and manage models without installing Python, Node.js, or Docker.
Use the built-in llama.cpp runtime, connect Ollama or LM Studio, or explicitly opt into Gemini 3.5 Flash. Local providers stay on the computer; cloud use is always a visible choice.
Alpha notice: the current installer is unsigned, so Windows SmartScreen may show an unknown-publisher warning. Verify the checksum on the release page and keep backups of important data.
Download Local Agent Studio 0.5.2 for Windows
Close any older Local Agent Studio window and tray icon, run the installer, then choose one of four guided AI providers. No separate Python, Node.js, or Docker installation is required for ordinary use.
Creating one useful agent should not require understanding runtimes, quantization, API servers, workflow schedulers, and half a dozen configuration formats first. Local Agent Studio turns that setup into a guided desktop experience while keeping the technical boundaries inspectable for advanced users.
- One Windows application: Electron, the backend, and the default runtime path are packaged together.
- Local AI without the maze: Hardware-aware guidance, model downloads, checksums, and safe defaults.
- Visual multi-agent workflows: Sequential and parallel agents can mix Gemini, llama.cpp, Ollama, and LM Studio in one workflow, alongside routers, reviews, functions, approvals, and outputs.
- Understandable execution: Watch each node move through pending, running, approval, completed, failed, or cancelled states.
- Human control: Consequential tools require approval tied to the exact action shown. Email asks by default and can be explicitly set to automatic per Send Email node; file writes, Python, MCP, and mutating HTTP cannot use that exception.
- Practical integrations: Attach documents or images to a run, configure email, create real
.docxand.xlsxfiles, and attach reusable skill files to individual agents. - Permission layers: Studio-wide master switches and per-agent permissions are both enforced by the backend for file, network, Python, MCP, and attachment access.
- Portable designs: Export
.agentpackworkflows without secrets, local paths, model files, or run history. - Honest provider choice: Gemini is convenient but not local; the app says so before a key is connected.
| Provider | Best for | Setup |
|---|---|---|
| Built-in llama.cpp | Maximum privacy and the fewest moving parts | Install the verified runtime and a GGUF model inside the app |
| Ollama | People who already use Ollama or want its model library | Open Ollama; the app detects it and can install the smallest starter model |
| LM Studio | People who manage models through the LM Studio desktop app | Start LM Studio's local API server and select a downloaded model |
| Gemini 3.5 Flash | Fast setup without downloading model weights | Create a Google AI Studio key through the guided opt-in flow |
The first-run wizard cannot finish until the selected provider has a runnable model. For local providers, Qwen 2.5 0.5B is offered as a small quick-start download.
Windows 10 22H2 or Windows 11 x64 is required. A signed installer is a release goal; current alpha installers are unsigned and may trigger Microsoft SmartScreen.
- Download the Windows installer.
- Open Local Agent Studio and review the private hardware assessment.
- Choose Gemini, llama.cpp, Ollama, or LM Studio.
- Connect/select at least one runnable model.
- Create the sample studio and run Research then write from the Workflows page.
Want a fully local setup? Choose built-in llama.cpp or an existing Ollama/LM Studio installation. Want the quickest cloud setup? Choose Gemini and follow the API-key guide. Agents using different providers can work together in the same visual workflow.
- Build private local AI agents with llama.cpp, Ollama, or LM Studio.
- Connect Gemini and local LLM agents in one multi-agent workflow.
- Create approval-aware email and document automation.
- Read, search, and generate Word, Excel, JSON, CSV, and text files.
- Prototype sequential agents, parallel branches, routers, reviewers, and human approvals.
- Share reusable
.agentpackworkflow designs without bundling secrets or model files.
For the full non-technical walkthrough, see the User Guide.
- Hardware, RAM, VRAM, storage, and provider detection.
- Verified llama.cpp runtime installation without administrator privileges.
- Resumable curated GGUF downloads with license acknowledgement and checksum support.
- Agent creation with bounded context, answer length, instructions, and local memory.
- Visual workflow creation with live per-node progress and cancellation.
- Meaningful Approval, Function, Router, Review, Parallel, Input, and Output nodes.
- Curated file, document, HTTP, transformation, and email tools plus human approval nodes.
- Encrypted Gmail, Outlook, Yahoo, and custom SMTP setup with a real test-send flow.
- Native Word and Excel reading, searching, and approval-gated file creation.
- Per-agent Markdown, text, JSON, and YAML skill attachments encrypted at rest.
- Run-time document/image upload and approved local-path input with strict size limits.
- Optional Python functions and local stdio MCP tools behind Developer Mode, explicit agent authorization, and exact-action approval.
- Opt-in grounded Google Search for Gemini agents when both studio and agent web permissions are enabled; local agents can use explicit allowlisted HTTP workflow functions.
- AES-256-GCM protection for sensitive values with a master key protected by Windows DPAPI.
- Authenticated loopback backend and inference endpoints bound to
127.0.0.1. - Optional encrypted Gemini API-key storage and cloud execution blocked by Offline mode.
- SQLite-backed projects, agents, workflows, benchmarks, schedules, and run history.
Local-first does not mean every possible action is automatically safe.
- llama.cpp, Ollama, and LM Studio inference stays on loopback by default.
- Gemini prompts leave the computer only for agents explicitly assigned to Gemini.
- Google controls Gemini free-tier quotas and data-use terms; the app links to the current terms before connection.
- External HTTP and email tools can send data outside the computer and are permission-gated.
- SMTP passwords and agent skill files are encrypted locally and excluded from
.agentpackexports. - Arbitrary Python and MCP execution are disabled by default. Developer Mode requires a studio switch, an agent permission, and approval for every call; it is powerful but is not an operating-system sandbox.
- The alpha installer is not yet code-signed and automatic updates are not yet implemented.
- Model weights have their own licenses; “runs locally” does not necessarily mean “open source” or “commercially unrestricted.”
Read SECURITY.md for trust boundaries, reporting, and release requirements, and DISCLAIMER.md for product-risk and backup guidance. Architecture details are in docs/ARCHITECTURE.md.
The first response includes model-loading time. A two-agent workflow invokes a model at least twice, so it is naturally slower than one chat response. The beginner defaults favor reliability on a 16 GB computer:
- 4,096-token working context;
- 128-token target response;
- 256-token hard output ceiling for ordinary agent nodes;
- one inference at a time when memory capacity is uncertain.
Use Resources → Private benchmarks to measure tokens per second and first-token latency on the actual computer. The 0.5B starter model proves the setup quickly; larger models usually produce better results.
Run monitoring uses a targeted lightweight endpoint rather than refreshing models and providers on every progress update. Hardware discovery is bounded and offers conservative defaults if Windows cannot report a device cleanly.
Contributor prerequisites:
- Windows 10 22H2 or Windows 11 x64
- Node.js 22+
- pnpm 11+
- Python 3.11 through 3.14
python -m venv .venv
.\.venv\Scripts\python.exe -m pip install -r backend\requirements.txt
pnpm install
$env:LOCAL_AGENT_PYTHON = (Resolve-Path .\.venv\Scripts\python.exe)
pnpm devClone the repository using GitHub's Code button, open PowerShell in that folder, and then run the commands above.
Validation:
pnpm typecheck
pnpm test
.\.venv\Scripts\python.exe -m pytest backend\tests -qAfter packaging, validate the real frozen backend and Electron interface:
.\scripts\smoke_packaged_backend.ps1
.\.venv\Scripts\python.exe .\scripts\audit_packaged_ui.pyBuild an unsigned local test installer:
$env:LOCAL_AGENT_PYTHON = (Resolve-Path .\.venv\Scripts\python.exe)
pnpm package:win:unsignedProduction releases use pnpm package:win with a Windows code-signing certificate. Credentials, model weights, local databases, and generated installers do not belong in Git.
backend/ FastAPI service, providers, scheduler, tools, storage, and tests
build/ Canonical application icon used by the UI, tray, EXE, and installer
docs/ Beginner guide, architecture, and screenshots
scripts/ Reproducible build, packaging, SBOM, icon, and QA helpers
src/main/ Electron lifecycle, backend process, tray, and security boundaries
src/renderer/ React interface, workflow canvas, onboarding, and tests
vendor/llama.cpp/ Runtime destination; binaries are downloaded, never committed
Focused issues and pull requests are welcome. Start with CONTRIBUTING.md, review the Roadmap, and keep privacy boundaries visible in every user flow.
Local Agent Studio is licensed under Apache-2.0. Third-party runtimes, applications, and model weights retain their own licenses; see THIRD_PARTY_NOTICES.md. The software is provided without warranty; read DISCLAIMER.md, keep independent backups, and obtain qualified legal review before commercial distribution.