Most AI tools today work the same way: you type a prompt, get a response, then manually do whatever the AI suggested. Lapu AI takes a fundamentally different approach. Instead of just answering questions, it executes tasks directly on your computer -- reading files, running commands, and orchestrating workflows across your desktop.
The agent loop#
When you give Lapu a task, it doesn't just generate a one-shot response. It enters an agent loop -- a cycle of reasoning, planning, acting, and observing that continues until the task is complete.
Here is what that looks like in practice. Say you ask Lapu to "refactor the authentication module to use JWT tokens." The agent will:
- Scan your project to find the existing auth files, understand the current implementation, and map out dependencies.
- Plan the changes by identifying which files need modification, what new code is required, and what tests should be updated.
- Execute step by step, modifying each file, running your test suite after each change, and adjusting course if something breaks.
- Report the results with a summary of every change made and the final state of your test suite.
At each step that involves writing or deleting files, running commands, or accessing sensitive resources, Lapu pauses and shows you exactly what it plans to do. You approve, modify, or skip each action.
How the agent reasons#
Not every task requires the same approach. A quick file rename is straightforward. A complex code refactor requires deep reasoning and careful planning. Lapu AI uses frontier language models to understand the nuance of each task and choose the right execution strategy.
The AI capabilities are built directly into the agent. You describe what you need, and Lapu figures out the best way to accomplish it. There are no API keys to manage, no model settings to configure, and no provider decisions to make.
Local execution#
Lapu AI executes tools on your machine. When it reads your files, those reads happen locally. When it runs a shell command, that command executes in your local terminal. There is no cloud workspace, no uploaded documents, no remote file storage.
When the agent needs to reason, relevant context is sent to AI model providers via Lapu AI infrastructure. No files are stored remotely — the model processes the context and returns results. This local-first architecture means there is no cloud workspace holding your data between sessions.
The permission system#
Trust is the hardest problem in AI agents. You want the agent to be powerful enough to be useful, but safe enough that it cannot cause damage. Lapu solves this with a granular permission system.
Every action is categorized by risk level. Low-risk actions like reading a file or listing directory contents can be auto-approved. Medium-risk actions like writing to a file require a single confirmation. High-risk actions like deleting files show a detailed preview and require explicit approval. Critical actions like running destructive shell commands require per-invocation confirmation and cannot be session-allowed.
You can customize these thresholds. Want to auto-approve all file reads in a specific project? Done. Want to require confirmation for every single action? That works too. The permission system adapts to your comfort level.
Real-world workflows#
Here are some examples of what Lapu AI can handle:
- Code reviews -- "Review the last 5 commits, check for security issues, and create a summary with recommendations."
- File organization -- "Sort my Downloads folder by file type, rename files to follow a consistent naming convention, and delete anything older than 30 days."
- Data processing -- "Read all CSV files in this folder, combine them into a single dataset, clean up the formatting, and generate a summary report."
- Project scaffolding -- "Create a new Next.js project with TypeScript, Tailwind, ESLint, and a CI pipeline. Set up the folder structure and add starter components."
Each of these tasks involves multiple steps, multiple tools, and decision-making along the way. That is exactly what Lapu's agent loop is built for.
Try it yourself#
The best way to understand how Lapu works is to try it. Download Lapu AI for macOS or Windows and start with a simple task right away. As you see the agent reason, plan, and execute, you will understand why working with an AI agent feels fundamentally different from chatting with a chatbot.
