Clear goals, appropriate permissions, and verifiable results are the beginning of stable use.
01
First time use
Configure the model in settings and pass the connection test. Select an empty project directory, first use chat mode to understand the problem, then use agent mode to perform operations.
Try describing the task like this
Create an HTML page that can add, complete, and delete to-do items. Save data to localStorage. When done, open the built-in browser, actually test these three operations, and provide the file link.
02
Models and local services
Cloud providers require your own key and model ID. Testing the connection makes a real request and may incur a small provider fee; whether a model supports tools is subject to the actual response.
Local model: Start the local API service of Ollama or LM Studio, then select the corresponding provider in BE AI and refresh the model list. The software will not automatically download weights, nor will it scan the disk to collect model files.
Image generation requires a separately configured supported image provider and model. Text-only providers such as DeepSeek will not appear in the image API selection.
03
Computer actions permission
System Settings → Privacy & Security → Screen & System Audio Recording / Accessibility: Check whether BE AI is authorized. After granting authorization, restart from the same installation location, then click the in-app permission check.
Grant these permissions only when computer operation is needed. Project access, computer access, and operation approval can be selected separately. Clicking Stop terminates the current execution and releases the mouse and keyboard.
04
Language and interface
Desktop: Settings → Language → select your preferred language. The official website uses the language menu in the navigation bar. The interface, system status, and prompts switch with the language; your own files, original chat text, code, and third-party web pages will not be translated without permission.
Arabic, Hebrew, and Persian use right-to-left layout. The model will answer according to the current language preference; you can still specify the output language in the task.
05
Long tasks and scheduled tasks
Steps per round are automatic checkpoints, not the task endpoint. The task will continue and be verified; if permissions, credentials, or persistent failures occur, progress is preserved and the real blocker is explained—no infinite retries.
Scheduled tasks run locally. Mac, BE AI, and required network services must remain available; after the computer is turned off, the official server will not execute local tasks on its behalf.
06
Encountered a problem
Submit feedback in the client or on the official website, clearly stating the version, reproduction steps, actual results, and expected results. Do not enter API Keys, passwords, or private files.
Feedback is saved like a conversation, and team replies can be viewed on the same device. To reduce abuse, each device and source IP can submit only one user feedback message per Beijing natural day in total; the same network may share the quota.
BE AI