Before you install anything
Choose a first task that is useful but reversible. Good examples include summarizing a folder of documents, analyzing a copy of a spreadsheet, comparing public sources, or drafting a report. Avoid giving a new agent authority to send messages, spend money, publish content, or modify your only copy of important files.
Make a backup and work from copies until you understand the system's permissions and output quality. Local execution improves control, but the agent can still misunderstand the task or use a tool incorrectly.
A six-step local workflow
Download the desktop app
Use the Local AGI download for your operating system. The website currently provides macOS and Windows builds.
Choose a bounded task
Specify the outcome, source material, time range, format, and what the agent must not do.
Limit the working set
Provide only the files or folders necessary for the task. Use copies for workflows that may edit data.
Select where AI runs
Use Local AI for work that should stay on your device. Choose a Private Network only when approved devices need to contribute. Use public capacity for suitable non-sensitive work.
Ask for an inspectable artifact
Request a report, table, saved file, or structured summary with source references—not merely a confident conclusion.
Review before reuse
Check calculations, citations, assumptions, and changes. Treat the output as prepared work, not an unquestionable decision.
Write a task the agent can verify
A strong task prompt defines five things:
- Outcome: what deliverable should exist when the task is done?
- Inputs: which files, URLs, dates, or folders are in scope?
- Constraints: what must remain unchanged or private?
- Evidence: how should claims and calculations be supported?
- Stop condition: when should the agent ask instead of continuing?
Analyze the attached sales CSV for the last 12 months.
Return a report with:
- monthly revenue and order trends
- the five largest changes
- possible explanations clearly labeled as hypotheses
- a table of the calculations used
Do not modify the source file. Ask before using any external service.Choose the privacy level task by task
Local AI
Use local execution when sensitive files should remain on one device or when you want direct control over the model and environment.
Private Network
Use a Private Network when your approved devices need to share intelligence. Communication between approved nodes is end-to-end encrypted; internet connectivity may still use a relay to cross network boundaries.
Public Network
Use broader capacity only for tasks whose data and tools are appropriate for public providers. Separate general research from confidential source material when possible.
Review checklist
- Did the agent use only the sources you authorized?
- Can you reproduce important calculations?
- Are factual claims linked to evidence?
- Are assumptions labeled instead of presented as facts?
- Did it alter any file or external system?
- Does the final artifact match the requested format?
Finance, health, legal, security, and current-event conclusions require qualified human review and current authoritative sources, regardless of where the agent runs.
What to try first
A private data-analysis workflow is a good first test because the inputs, calculations, and final report can all be inspected. Before choosing a model, read the difference between a local model and a local agent.
Start with one bounded task
Download the app, keep the first working set small, and ask for a result you can independently check.