Finalist
DocRight Context
Coding agents like Claude Code, Cursor, and Copilot write great code but hallucinate APIs.
Core trick
Agent reads llms.txt indexes, selects relevant doc pages across sources, and fetches only those before generating code.
Description
Coding agents like Claude Code, Cursor, and Copilot write great code but hallucinate APIs. Ask one to "build a CrewAI agent with Composio Slack tools" and it uses method names that don't exist. Why? Stale training data, not actual docs.
DocAgent fixes this. It indexes doc sites via llms.txt, reasons about which pages are relevant across multiple sources, fetches only those, and generates code using real API patterns.
Today: your coding agent guesses, you debug. With DocAgent: it reads the real docs first, your code works.
Long term, any AI tool calls our API for accurate context before generating code. Via Skyfire, agents pay per query to learn new APIs. As doc sites adopt llms.txt, DocAgent scales with zero re-engineering.
Your coding agent hallucinates APIs. Ours reads the docs first.
README excerpt
# llama-lounge-agentic-HACKATHON # DocAgent — Architecture & Approach ## Problem Statement Coding agents like Claude Code, Cursor, Windsurf, and Copilot write great code but hallucinate APIs. Ask one to "build a CrewAI agent with Composio Slack tools" and it uses method names that don't exist, import paths that changed months ago, and authentication flows it invented. Why? Stale training data, not actual docs. No existing AI tool can autonomously navigate multiple documentation sources, cross-reference them, and generate verified code. **DocAgent is the missing layer between coding agents and documentation.** Given a task, it autonomously indexes doc sites via `llms.txt`, reasons about which pages across multiple sources are relevant, fetches only those pages, generates code using real API patterns, and validates the output in a sandbox. It reads the actual docs so your AI doesn't have to hallucinate. ## Our Solution **DocAgent** is a multi-agent system (CrewAI + Composio) that acts like a senior developer who has read all the docs. You give it a task and a set of documentation sources. It: 1. **Indexes** each doc site via `llms.txt` (lightweight table of contents, not full content) 2. **Navigates** using an AI agent that reasons across ALL indexed sites to pick relevant pages 3. **Fetches** only the right pages from the right sources as clean markdown 4. **Generates** complete, production-ready code using a multi-agent crew 5. **Validates** the generated code via loc