CocoIndex is an incremental data framework for AI workloads: you declare the target state you want — an index, a knowledge graph, embeddings in a store — and its Rust-powered engine continuously computes only the deltas needed to keep that target fresh as sources change. It has earned about 11,700 stars on GitHub and is in very active development in 2026.
Agent and RAG pipelines quietly rot when their data goes stale, and recomputing everything on every change is too slow and expensive. CocoIndex borrows React’s mental model for data engineering: you write a straightforward transformation in Python as if it were a one-off job, and the engine derives the dependency graph, tracks lineage for every byte, and thereafter re-runs only what a source edit, a schema change or a code change actually affects — with automatic schema evolution and no downtime or index swaps. Its flagship application, CocoIndex Code, builds AST-based coding context for agents: call graphs, symbol tables and semantic indexes over a repository that stay sub-second fresh as code changes, so coding and review agents reason over the current codebase instead of a stale snapshot. The same engine indexes documents, meeting notes and other sources into stores such as Postgres for retrieval. It is production infrastructure for the unglamorous problem underneath every long-horizon agent: keeping its view of the world true.
CocoIndex is free and open source under the Apache-2.0 license, and its website lists no paid tier. You run it on your own infrastructure and pay only for your compute and storage.
CocoIndex is for data and AI engineers building RAG pipelines, codebase intelligence and agent context systems whose value depends on freshness — teams that cannot afford agents reasoning over stale indexes or full recomputes on every change.