Overview

Two products share the name. The open-source framework is MIT licensed and gives you loaders, chunkers, retrievers and, since June 2026, the Workflows 1.0 event-driven agent runtime in both Python and TypeScript. LlamaCloud is the paid half: Parse, Extract, Classify, Split and Index, all drawing on one credit balance at $1.25 per thousand.

The honest pitch is document ingestion. Retrieval quality is usually bounded by whether the parser understood the table on page 41, not by which ANN index sits underneath, and the LlamaParse tiers let you spend selectively: 1 credit a page for fast, 3 for cost-effective, 10 for agentic, 45 for agentic plus. At 10 credits that is roughly $0.0125 a page, defensible for contracts and regulatory filings, much harder to defend across a million pages of clean HTML you could have parsed for nothing.

Buy it when your corpus is genuinely difficult: nested tables, scanned forms, mixed-language filings, the material where a generic extractor loses half the meaning. Be more cautious about adopting the framework wholesale. It moves fast and it deprecates, llama-deploy to llama-agents being only the latest instance, and teams that wrapped their entire retrieval stack in its abstractions have had to unpick that work later. A thin layer over code you own ages better.

Key Features

✓ Workflows 1.0, released 22 June 2026, an event-driven agent runtime in Python and TypeScript with its own repository and package

✓ LlamaParse v2 parsing tiers priced per page at 1, 3, 10 and 45 credits, so you spend more compute only on documents that need it

✓ LlamaCloud modules for Parse, Extract, Classify, Split and Index, all metered from a single credit balance

✓ Hundreds of loaders and vector-store integrations, including Qdrant, pgvector, Pinecone and Elastic

✓ llama-agents-server, which exposes a workflow as a REST API and replaces the deprecated llama-deploy

My Take

Version churn has burned people here. llama-deploy is deprecated, the query-engine abstractions of two years ago have been quietly superseded and a good deal of published tutorial code no longer runs, so treat anything older than about a year as archaeology. What deserves attention is LlamaParse and the extraction layer, because ingestion is where most RAG systems fail long before retrieval gets a chance to. Pricing moved against buyers this year: credits went from $1 to $1.25 per thousand, and Agentic Plus parsing at 45 credits a page lands near $0.056, a number Reducto and LandingAI's ADE will happily contest on either cost or accuracy depending on what your documents look like. Workflows 1.0, shipped 22 June 2026, is the cleanest thing the project has produced.