vtx-coding-agent

Vtx Coding Agent is designed as a minimalist, modular harness for AI-assisted coding, prioritizing transparency and efficiency. Unlike many agents that consume significant context with hidden boilerplate, Vtx maintains a lean runtime footprint of approximately 2,600 tokens (o200k_base encoding), ensuring that the majority of the model's context window remains available for user code and conversation. This design choice aims to reduce costs and latency while providing an auditable system prompt. The project offers multiple interaction surfaces: a keyboard-driven Terminal User Interface (TUI) built with Textual, a headless CLI for non-interactive scripts and CI pipelines, and a Python SDK for programmatic agent construction. It supports over 50 LLM providers, including OpenAI, Anthropic, Azure, DeepSeek, and local models via Ollama or vLLM. Users can also configure custom OpenAI/Anthropic-compatible endpoints via YAML. Key features include a set of 10 default tools (read, edit, write, bash, find, skill, web, ask_user, task, goal) and dynamic context loading from AGENTS.md or CLAUDE.md files. Vtx supports 'handoff agents'—named profiles with specific instructions and tool permissions that can be cycled live—and task sub-agents for delegating isolated work. Persistent goals allow for durable, file-backed objectives with auto-continue checkpoints and completion audits. Safety and extensibility are central to the design. The default 'prompt' permission mode gates mutating tools, blocking destructive commands unless explicitly allowed. The system is self-extensible, allowing users to add Python tools, intercept calls, or define lifecycle hooks via YAML. An extension manager facilitates installing packages from PyPI or GitHub.