The archive · Developer & Business Tools · Product decision · 2026
LiteParse: LlamaIndex bets AI agents will parse documents locally, not in the cloud
LlamaIndex (run-llama) open-sourced LiteParse, a Rust document parser with no cloud or LLM calls; the repo drew 12,250 GitHub stars by September 2026.
LlamaIndex (run-llama)
What the business is
LiteParse is an open-source, local-first document parsing library from LlamaIndex (run-llama): a Rust core that extracts spatial text, bounding boxes and Markdown from PDFs, Office files and images, with built-in OCR and no cloud dependency, API key or LLM call.
How it started
LiteParse appeared on GitHub in February 2026 as run-llama's standalone open-source parser focused 'exclusively on fast and light parsing', without proprietary LLM features or cloud dependencies. The docs name real-time applications, coding agents and local workflows as its design targets — the cases where a cloud round-trip is too slow, too costly or too privacy-sensitive.
What happened
By the 2026-09-04 crawl the repo had 12,250 stars, and the product had grown from a CLI into Rust, Node/TypeScript, Python and browser WASM bindings, a complexity check that routes scanned or multi-column documents before a full parse, and an agent-skill wrapper for Claude Code, Cursor and Codex. The README states the open-core split openly: LiteParse for fast local parsing, and LlamaParse, 'our cloud-based document parser', for dense tables, multi-column layouts, charts, handwritten text and scanned PDFs.
No ending yet — it is still running.
Background
LiteParse is an open-source document parser from LlamaIndex (run-llama), the company behind the LlamaIndex framework and the cloud service LlamaParse. Written in Rust, it parses PDFs, Office files and images entirely on the user's machine, returning spatial text, bounding boxes and clean Markdown, with built-in Tesseract OCR and no cloud dependency, LLM call or API key.
The bet is that the default document layer for AI agents should be local, fast and free. The docs name real-time applications, coding agents and local workflows as the design targets, and the README calls LiteParse a standalone open-source tool 'without proprietary LLM features or cloud dependencies' while steering genuinely hard documents — dense tables, scans and handwriting — to LlamaParse, the company's cloud parser.
The repo appeared on GitHub on 2026-02-09 and carried 12,250 stars at the 2026-09-04 crawl, with bindings for Rust, Node/TypeScript, Python and browser WASM plus an agent skill for Claude Code, Cursor and Codex. The open-core split is the product strategy: LiteParse wins the default install, and LlamaParse wins the documents too hard for local parsing.
What has to be true
- Local-first removes the two objections that stop agents from parsing documents at scale: per-page cloud cost, and sending private or regulated files to a third-party parser.
- Fast and light is a real technical bet: PDFium-based spatial parsing plus a cheap complexity check fits real-time and agent-loop use cases that a heavyweight LLM parser cannot serve.
- Open-sourcing the base layer buys distribution — 12,250 GitHub stars in seven months — and every LiteParse integration is a potential upgrade path to cloud LlamaParse.
What can be applied
Open-core works when the free tier is the right tool for the common case: give away a fast local parser as the default, and the paid cloud survives only for documents it cannot handle.
Aftermath
As of 2026-09-05 LiteParse is live and actively documented: the README points users of an earlier LiteParse V1 to separate 'old code', the docs cover CLI and library use across four language stacks, and the site keeps the split explicit — local parsing as the default, LlamaParse as the answer for documents too hard for local tools. No funding or revenue figures appear in the material; the bet is still being tested.
Sources
spotted an error? The archive wants to know.
Your turn
You just read one. Describe what you are building, and see who is betting on the same thing.
Free account · 3 free questions · no card