MultiAgentOS
by lukekevinmclaughlin-oss
MultiAgentOS is a local-first desktop workspace where specialised agents plan, code, browse, automate, and operate — across 162 built-in tools — with your own local or cloud LLMs. Full IDE, real desktop control, and an in-app self-test that proves every tool works on your machine.
MultiAgentOS is a full desktop application — agent rail, chat workspace, live run telemetry, built-in code editor, knowledge base, and granular privacy controls. Below is the actual app.
Coder, Operator, Planner, and Reviewer live in the left rail, each with its own assigned LLM and idle/working state. The chat surface in the middle is where you talk to the team — pick a mode (Chat, Code, Vibe Build, Operator, Auto Agent, Broadcast) and a Guardrails preset. The right panel toggles Timeline, Files, Code, Terminal, Models, and Auto.
The Timeline panel streams a live, append-only log of every tool the active agent calls during a run — lease issued, browser.search, file.write, verifier checks, lease released. Nothing is hidden. You can scrub back through prior steps at any point.
The Code tab is a full file browser plus editor. Switch between
per-project Workspaces and the System scope. Open, create, rename, or
delete files; create new workspace files inline. Agents reach into the
same tree via file.* tools — what you see in the editor is
what they see.
MultiAgentOS ships with a queryable knowledge base of functional specs — load a single LLM, prompt a loaded LLM, broadcast across models, switch active LLM, unload, and many more. Agents look these up to answer “does the software support X?” without you having to teach them.
Flip on Offline Mode and MultiAgentOS will only use local providers and models — every cloud provider (OpenAI, Anthropic, etc.) is disabled. The Provider Health panel shows which providers are reachable right now. No telemetry. No background uploads. No surprises.
Every prompt routes through a mode. Switch any time without losing context. Pair any mode with a Guardrails preset (Full Access, Ask Each Time, Minimal, Sandbox) to set how much the software will let the LLM do.
Plain conversation with the active agent. No tool calls unless asked. Best for thinking out loud, drafting, or asking for explanations.
Coder takes the wheel. File reads, file writes, shell, git, python.exec — all wired in. Diffs surface in the right panel.
Conversational scaffolding — describe what you want and the team plans, scaffolds, builds, and tests. Best for new features or fresh projects.
Browser and desktop control. The Operator agent searches the web, opens apps, fills forms, takes screenshots, and clicks for you.
Autonomous mode. The team runs without per-step approval — gated by your Guardrails preset and lease policy. For longer, multi-step tasks.
One prompt, every loaded LLM. See the team's answers side-by-side; the orchestrator stitches a consensus answer with attribution.
MultiAgentOS ships with a full agentic toolbelt. Web, browser, desktop, filesystem, shell, git, documents, data, image, voice, calendar, memory, and more — all wired into a unified runtime your agents pick from.
Search, fetch, scrape, summarise. Live weather, stock, crypto, news, sports — public-endpoint by default.
Visible search and URL opens for the user, plus headless Playwright for scripted scraping, clicks, screenshots.
Read and write across PDF, DOCX, XLSX, PPTX. Compare files, search content, copy/move/delete with workspace safety.
Run Python in a sandboxed subprocess. Read/write CSV. Render charts to PNG via matplotlib.
Real OS-level control via a native Tauri bridge. Open apps, take screenshots, click, type, hotkeys.
Inspect any local repo. Status, diff, log, blame, branch, search commit history.
Search repos, read PRs and issues, read files at any ref. Uses your gh CLI auth or public REST.
Resize, thumbnail, convert formats, read metadata via Pillow. Generate via OpenAI / Stability when configured.
Speak text out loud. Transcribe audio via local whisper.cpp, OpenAI-Whisper Python, or hosted Whisper.
Persist key/value memory across runs. Recall, search, forget — fully local.
Create .ics events importable into any calendar app. Local reminders. Windows scheduled tasks.
Gmail, Google Calendar / Drive / Docs / Sheets, Slack, SharePoint — all gated behind your own credentials.
MultiAgentOS ships with four built-in agents — Planner, Coder, Reviewer, Operator — each with its own role, system prompt, and assignable LLM. Add custom agents for domain work. The orchestrator coordinates handoffs and collaborative broadcast so multiple loaded models can work in parallel.
Breaks complex requests into ordered steps and assigns work to the right agent. Tool families: llm.*, file.*, git.*.
Reads files, writes code, runs tests, applies patches. Tool families: file.*, shell.*, git.*, python.exec, data.*.
Reviews diffs, checks for bugs and style issues, reports concrete file:line findings. Tool families: git.*, file.*, shell.*.
Drives browser and desktop. Searches the web, opens apps, fills forms, takes screenshots. Tool families: browser.*, desktop.*, native.*, shell.*.
MultiAgentOS uses an explicit tool-authority lease model. The software owns full tool access. When an LLM activates on a run, the backend issues a run-scoped lease that transfers that authority for the duration of the task. When the task ends, the lease is released.
Secure Stripe checkout · instant licence delivery · 14-day refund.
Yes — bring your own. MultiAgentOS supports Ollama, LM Studio, llama.cpp,
local GGUF files, EXL2, MLX, plus OpenAI / Anthropic / Google / Groq /
Together / Mistral via API key. Recommended local model for tool-calling
self-test: qwen2.5:7b-instruct.
MultiAgentOS includes a built-in diagnostic that runs every registered tool against safe inputs and reports which work on your machine. Useful for confirming a fresh install, debugging a missing dep, or proving the tool surface to your team.
A single licence is for one person to install on up to two devices they own — a desktop and a laptop, for example. The licence is yours; it does not expire.
14-day money-back guarantee. Email support@multiagentai.ai within 14 days of purchase and we'll refund you in full — no questions, no forms. The licence is then deactivated on every device.
Almost none, by default. Activation pings the licence server with a stable device ID and your licence key. That's it. Project files, prompts, and agent traces stay local unless you explicitly route work through a cloud provider you configure.
Because the work happens on your machine, not on our servers. There's no inference bill we have to amortise across customers. You pay for the version you bought; future major versions are separate purchases.