CocoIndex enables semantic code search
- CocoIndex’s open-source code indexing stack is being adopted as a way to search repositories by intent, not filenames, giving coding agents fresher context. - The company’s `cocoindex-code` tool says it uses abstract syntax tree parsing, works with Claude, Codex and Cursor, and can cut tokens by 70%. - Sourcegraph and CocoIndex both frame semantic search as a layer on top of keyword and structural search, not a replacement. (sourcegraph.com)
Semantic code search means looking for code by what it does, not by the exact words in a filename, function name or comment. (sourcegraph.com) The basic trick is to turn code and a natural-language query into numerical representations, then compare them for conceptual similarity. Sourcegraph says that lets a search for “retry logic for failed API calls” surface files that do not contain the word “retry.” (sourcegraph.com) CocoIndex is one of the open-source projects pushing that approach into developer tools. Its site says it keeps a repository index in sync with every commit and feeds coding or code-review agents “structure, not raw text.” (cocoindex.io) The company released `cocoindex` 1.0.0 on PyPI on April 21, 2026. Its package description says the system maintains derived indexes incrementally, recomputing only changed data instead of rebuilding everything. (pypi.org) A companion repository, `cocoindex-code`, packages that idea as an embedded code-search engine. Its GitHub README says it is abstract-syntax-tree based, works from the command line, and can integrate with Claude, Codex, Cursor and other agents through Skill or Model Context Protocol. (github.com) Abstract syntax trees are the parsed map of a program’s structure — functions, classes, calls and blocks — rather than raw lines of text. CocoIndex says that structure helps an agent search for meaningful units of code instead of shoveling full files into a prompt. (github.com) (cocoindex.io) That matters because large codebases break simple grep-style workflows. Sourcegraph’s documentation says natural-language search works best when paired with its existing code search and code navigation tools, which the agent uses in a loop to refine an answer. (sourcegraph.com) Sourcegraph’s Deep Search is a close commercial example of the same direction. The company says the tool answers questions about a codebase by combining multiple search modes, following references, reading files and returning a sourced explanation. (sourcegraph.com 1) (sourcegraph.com 2) The practical pitch is not that semantic search replaces exact matching. It is that developers and coding agents can start with meaning, then narrow with path filters, references and exact-text search when they need exhaustive results. (sourcegraph.com 1) (sourcegraph.com 2) CocoIndex’s own marketing language leans into that same use case: “Less grep. Find by meaning.” In the current wave of coding agents, the fight is increasingly over which tool can hand the model the right files before it starts writing code. (cocoindex.io)
