Skip to content

configuration.md

Latest commit

 

History

History
108 lines (77 loc) · 5.36 KB

configuration.md

File metadata and controls

108 lines (77 loc) · 5.36 KB

Configuration Reference

All Exe Computer Use settings are accessible through the Settings panel in the application UI. This document provides a complete reference of every configurable option.

Table of Contents

VLM Provider Settings

Configure the Vision Language Model endpoint used for action prediction.

Setting Description Default Notes
vlmProvider Model provider type -- Select from available providers in the dropdown
vlmBaseUrl API base URL -- Must be an OpenAI-compatible endpoint (e.g., https://api.openai.com/v1)
vlmApiKey API key -- Authentication key for the provider. Stored locally, never transmitted except to the configured endpoint
vlmModelName Model name ui-tars The model identifier sent in API requests
useResponsesApi Use OpenAI Responses API false When enabled, uses OpenAI's Responses API instead of Chat Completions. Enables previousResponseId for conversation continuity

Provider Examples

OpenAI-compatible endpoint:

  • Base URL: https://api.openai.com/v1
  • Model name: ui-tars (or your deployed model name)

Local llama-server:

  • Base URL: http://localhost:11435/v1
  • API key: (leave empty or use any string)
  • Model name: ui-tars

Local Model Settings

Configure the built-in llama-server for running UI-TARS models locally on your machine.

Setting Description Default Notes
localModelEnabled Enable local model serving false When enabled, the app manages llama-server processes
localModelAutoStart Auto-start servers on app launch true Servers start automatically when the app opens
localModelMainPort Main VLM server port 11435 Port for the primary action-prediction model
localModelReflectionPort Reflection server port 11436 Port for the reflection model used by RMA

Supported Models

Model Size Purpose Recommended For
UI-TARS-2B ~2 GB Action prediction (screenshot to action) Machines with 16 GB RAM
UI-TARS-7B-DPO ~7 GB Reflection and self-correction Machines with 32 GB RAM

When local model serving is enabled:

  1. The app downloads the llama-server binary if not already present.
  2. Model weights are downloaded from HuggingFace on first use.
  3. Each model runs as a separate llama-server child process.
  4. Health checks run periodically to verify server readiness.

Agent Settings

Control the behavior of the GUIAgent automation loop.

Setting Description Default Notes
maxLoopCount Maximum agent loop iterations 100 Safety limit to prevent infinite loops. The agent stops after this many screenshot-action cycles
loopIntervalInMs Delay between iterations (ms) -- Optional throttle between loop cycles. Useful for slowing down the agent for observation
rmaEnabled Enable Reflection Memory Agent true Enables automatic loop detection and self-correction. See Architecture: RMA
operator Operator type LocalComputer Which platform operator to use for action execution

Operator Types

Value Description
LocalComputer Controls the local desktop via nut-js (mouse, keyboard, hotkeys)
LocalBrowser Controls a web browser via Puppeteer

Reflection Settings

Configure the Reflection Memory Agent's model endpoint. These settings are separate from the primary VLM settings, allowing you to run a different model (or the same model on a different port) for reflection.

Setting Description Default Notes
reflectionBaseUrl Reflection model API URL -- OpenAI-compatible endpoint for the reflection model. When using local models, this defaults to http://localhost:11436/v1
reflectionModelName Reflection model name ui-tars-7b-dpo The model used for self-correction when a loop is detected

How Reflection Works

When the Reflection Memory Agent detects the agent is stuck (repeated similar screenshots), it:

  1. Gathers recent action history from the knowledge base.
  2. Sends a reflection query to the configured reflection model.
  3. Receives corrective guidance.
  4. Injects the guidance into the agent's next prompt to steer it toward a different approach.

For details, see Architecture: Reflection Memory Agent.

Environment Variables

These environment variables are used during build and development. They are not required for normal usage.

Variable Description Context
EXE_APP_PRIVATE_KEY_BASE64 Base64-encoded private key for app signing Build-time only. Used by electron-forge for code signing distributable builds
NODE_ENV Node environment Set to production automatically during builds. Controls optimizations and debug behavior
CI Continuous integration flag Set to e2e for E2E test builds to adjust packaging behavior