LlamaIndex releases LiteParse for 50+ document types

• LlamaIndex has open-sourced LiteParse as a “model-free document parsing tool” for AI agents, with the launch thread saying it runs locally, is free, and handles 50+ file formats without GPUs.
• The core pitch is speed plus layout preservation: the launch thread claims LiteParse can process “~500 pages in 2 seconds” on commodity hardware and keep complex tables in a spatial grid instead of flattening them into a text stream.
• LlamaIndex is positioning it as an agent-native tool rather than just another parser, with the agent install post showing a one-command skills install and integrations across Claude Code, Cursor, Warp, OpenClaw, and other agents.
• The company is also drawing a boundary around where it fits: the launch thread says LiteParse is “not a replacement for a VLM-based OCR tool,” making it a fast first pass for ordinary docs rather than the answer for the hardest scans and layouts.

LiteParse is a new open-source CLI and library for local document parsing. In the product repost, LlamaIndex describes it as a “lightweight and open source CLI and TS library,” which matters for engineers who want to wire parsing into TypeScript-heavy agent stacks instead of treating it as a hosted API.

The technical hook is that LiteParse tries to preserve document structure without using models. The launch thread says it is “more accurate than PyPDF, PyMuPDF, MarkItDown” on readability and shows table output as a spatial grid; the accompanying table parsing demo contrasts that with a competing parser that emits a messy sequential list. LlamaIndex’s GitHub repo also frames it as layout-aware extraction with OCR and screenshot support for cases where text alone is not enough.

A separate open-source repost summarizes the tool as “lightweight, local” with “no API calls,” reinforcing that the release is aimed at teams that need predictable local execution rather than another remote parsing dependency.

LlamaIndex is packaging LiteParse directly into agent workflows instead of asking developers to build glue code first. In the install post, the setup is a single npx skills add command, and the company says the parser can plug into “46+ different agents,” including Claude Code, Cursor, Warp, and OpenClaw.

That packaging matters because the parser is meant to do two jobs in coding agents: solve document-reading tasks directly, and turn PDFs, office files, or images into context that the agent can use while writing code. The Claude Code demo shows the install flow and a document-analysis task inside Claude Code, while the launch thread adds that LiteParse supports native screenshotting capabilities. The broader positioning in LlamaIndex’s blog post is clear: use LiteParse for the fast, cheap first pass, and escalate to heavier VLM-based parsing only when the document is too complex for text parsing alone.