MultiAgentOS
Beginner guide
MultiAgentOS is a desktop AI agent. That means it is not just a chatbot: it can read context you provide, use tools you allow, work in a browser, interact with a visible desktop workspace, run commands, and coordinate subagents. This guide starts from zero and shows a practical, safe workflow.
Step 1
The best first session is deliberately small. You want to learn where the controls are before you ask the agent to touch important files or accounts.
Use the Mac installer on macOS 14 or newer, or the Windows installer on Windows 10 or 11. Open the app, allow normal operating system permissions, and keep your first run focused on reading or planning rather than changing things.
The chat is where you describe the task. The context area is where files, folders, screenshots, and URLs enter the conversation. The sidecars show visible browser or desktop work. Settings control models, tools, prompts, and permissions.
Try: "Read this file and summarize the main decisions in plain English. Do not edit anything." This teaches you how attachments, instructions, and review work without risk.
Step 2
MultiAgentOS can work with hosted model APIs or local model servers. Use the simplest reliable option first, then tune privacy, cost, and speed as you learn.
Choose this when you want strong hosted models from a provider. Add the provider, model name, and API key in settings. This is usually easiest for complex reasoning, research, writing, coding, and multi-step tasks.
Choose this when you run Ollama, LM Studio, llama.cpp, or another model server on your own computer. Add the local endpoint, select the model, test the connection, and use smaller tasks while you learn the model's limits.
CLI pipe, OAuth, terminal, and local AI modes are useful when a provider or internal workflow exposes models differently. If you are new, only use these after the basic chat workflow feels comfortable.
Step 3
Good results usually come from three things: the desired output, the facts the agent should use, and the actions it is allowed to take.
Instead of "help with sales", write: "Review these three call notes and draft a follow-up email for each lead. Keep the tone friendly and do not invent product details." Specific instructions reduce wandering.
Attach the files, folders, screenshots, or URLs the agent should use. For local work, include the actual project folder or document. For web work, include the target site and login-state assumptions.
Tell the agent when to pause: before sending messages, deleting files, purchasing anything, changing account settings, running destructive commands, or submitting forms. The pause rule is your seatbelt.
Step 4
Tools make an AI agent useful, but every tool expands what it can affect. Start narrow and add capability as the job demands it.
Use file access for summaries, rewrites, code review, document cleanup, and project analysis. Ask for a plan first, then approve edits after you understand what will change.
Use browser actions for research, form filling, account navigation, and comparison tasks. Use desktop actions when the task lives inside a local app. Watch the sidecar so the work remains visible.
Use shell commands for development, diagnostics, file conversion, and automation. Add MCP servers or custom tools when you need structured access to databases, APIs, issue trackers, docs, or internal systems.
Step 5
Subagents are best for tasks that can be split into separate lanes, such as research, test writing, content drafting, data cleanup, or comparison work.
One subagent can inspect docs, another can test code, and another can draft a summary. Avoid giving every subagent every tool. Clear roles make review easier.
Use turn budgets, tool allow-lists, and short deliverables. Ask each subagent to report evidence, files read, assumptions, and anything it did not verify.
Treat subagent output as a strong draft, not magic. Check claims, inspect changed files, run tests, and decide which recommendations should become final work.
Safety workflow
Use this checklist whenever the agent can change files, operate a browser, run commands, or interact with another app.
Back up important files, attach only relevant context, state what success looks like, and name actions that require your approval. Keep sensitive credentials out of prompts unless absolutely necessary.
Watch visible actions, interrupt if the agent drifts, and ask it to explain the next step before risky moves. For coding work, ask for small patches and tests rather than one giant edit.
Review changed files, command output, browser submissions, account settings, and generated documents. Ask for a final list of what changed and what still needs human judgment.
Example workflows
These tasks are useful, realistic, and easy to supervise while you build confidence.
Attach a folder of PDFs, notes, or meeting transcripts. Ask for a one-page summary, an action list, and unanswered questions. Keep editing disabled for the first pass.
Give the agent a research question and let it use the browser. Ask it to collect links, compare sources, and separate verified facts from assumptions before writing an answer.
Attach the file, describe the desired columns, and ask for a preview of planned cleanup steps. Approve edits only after you understand how rows will be changed.
Attach your notes and ask for drafts, not automatic sending. Tell the agent the tone, recipient, and forbidden claims. Review every message yourself.
Attach the repo and ask for a map of the project, risky areas, and a first test command. Let the agent run read-only diagnostics before it edits code.
Troubleshooting
Most agent problems are fixable by narrowing the task, adding context, choosing a stronger model, or reducing tool access.
Add source files, paste examples of the output you want, and ask the agent to cite which file or page supports each claim.
Ask it to stop and explain the blocker. Then provide missing context, switch model, reduce the task, or allow one extra tool at a time.
Give it a smaller goal, require a written plan before actions, and set a strict pause rule before edits, browser submissions, or shell commands.