XAgent

XAgent is an experimental self-hosted AI agent for multi-step coding, research and data-analysis tasks. It plans subtasks, revises plans as work progresses and dispatches actions to agents. A command-line interface and browser UI let you guide a run or provide missing information.

Its Docker-based ToolServer supplies a file editor, Python notebook, web browser and Bash shell. Agents can work with files, execute analysis code, plot results and call external APIs through Rapid API. Generated files and execution records stay in a local workspace for inspection and replay.

The stock OpenAI configuration selects GPT-4-32K, which has been retired. Its model-name allowlist does not accept arbitrary contemporary OpenAI models, so replacing that default is not simply a model-setting change.

A separate documented local route uses XAgentGen to serve XAgentLlama through a CUDA-enabled Docker container, then launches XAgent with assets/xagentllama.yml. The 7B and 34B preview weights and the amd64 server image remain obtainable. These are legacy research distributions from 2023, not current frontier models; the local route requires NVIDIA GPU resources and model downloads. External website and API tools still need network access.

The application code uses Apache 2.0; the model weights carry separate Llama 2 terms. Developers can add tools and agents, but task completion and generated code need review.