CocoIndex is an open-source Python framework that keeps code, documents, and meeting notes ready for agents by reprocessing only what changed. You declare the target in Python. The engine memoizes results when the inputs and the function code are unchanged, tracks which source items and which logic are affected, and runs independent work in parallel. Install the framework with pip install -U cocoindex. The same page describes CocoIndex Code Plus as a licensed code-search product, built on CocoIndex Plus, that searches indexed repositories and follows symbol references.
for People building agents and LLM apps that need a current view of code, documents, or meeting notes. The page's examples are a repo index for code-review and coding agents, a meeting-notes knowledge graph, and a repository wiki. A Claude skill on the same page is for writing the pipeline. CocoIndex Plus is the path the page points to for running at scale.
use this if you want a coding agent to search a repo by meaning and get back small chunks, or you want an index of code and documents that only reprocesses what changed.
open cocoindex.io ↗Features
$ features --list
- ›cocoindex-code is an Apache-2.0 CLI (ccc) and a stdio MCP server. It parses files with Tree-sitter, tries to split them at function and class boundaries, and embeds the chunks. Chunks target about 1,000 characters, which the page calls about 300 tokens
- ›The MCP search tool defaults to 5 results (limit 1 to 100). The page says the code in that result is at most about 1,500 tokens, plus file paths, line numbers, and scores. A whole-file read costs the length of the file, and that file is sent again on later turns until the client compacts the conversation
- ›ccc search, which the skill runs, returns up to 10 results by default. --limit 5 matches the MCP default. offset, a path glob, and a language filter narrow the call. The MCP parameter names are limit, offset, paths, and languages
- ›Install with pipx install 'cocoindex-code[full]' or uv tool install --upgrade 'cocoindex-code[full]', then ccc index in the repo. Claude Code: npx skills add cocoindex-io/cocoindex-code, or claude mcp add cocoindex-code -- ccc mcp. Codex: codex mcp add cocoindex-code -- ccc mcp. Other stdio clients start ccc mcp with the project as the working directory
- ›The MCP search tool refreshes the index before each query by default. ccc index re-indexes changed files. With the local model, indexing cost is compute on your machine. With a cloud embedding provider, you pay that provider per token
- ›The page says the saving depends on the repository and the task, and to measure it by running the same task with and without the search tool. Search does not change how many tokens the agent writes. Questions about callers and references need Serena or CocoIndex Code Plus (ccx), which is a separate self-hosted product
Pricing
$ pricing --tiers
Open source
Free
pip install cocoindex. One flow process, embedded LMDB state, a local target. The page says LMDB has one writer, and local storage and write throughput can limit how large a deployment gets. The same flow code runs on Plus with a different SDK package and state store. The comparison marks the GitHub connector, multi-node compute, and shared components as not included here.
CocoIndex Plus
Quote
pip install cocoindex-plus, plus a license key. Shared Postgres state on your cloud or your own servers, including Kubernetes and a Helm chart. The page does not list a dollar price. It says they answer with a deployment plan and a license. Multi-node compute, sharding, and dynamic sources are marked on request. The usage dashboard is marked in development. OpenTelemetry export is marked available.
Why this rank
$ diff --rank 1
Ranked #1 in Context & Indexing. The job is handing an agent the right code and docs without paying to reread the repo. A default cocoindex-code MCP search returns at most about 1,500 tokens of code, five chunks of about 300 tokens, instead of whole files, and a later index run skips files that did not change. The same page says there is no single saving that holds for every repo, and that callers and references need Serena or the licensed Code Plus product. Serena ranks second as that symbol tool. Sourcegraph ranks third as the enterprise code-search instance. Repomix ranks fourth as a one-shot pack of the repo into a file. This order follows the job, not GitHub stars. The scores are unmeasured placeholders, and the site does not display them.
Strengths
$ strengths
- ›Apache-2.0. On 2026-11-11 the GitHub API reported 11,875 stars and 931 forks. The repository was created 2025-03-03. GitHub's language breakdown is led by Rust, then Python. The README says the core is Rust, with a Python SDK you install using pip install -U cocoindex
- ›Latest GitHub release on that date was v1.2.0, published 2026-11-09. It adds a Discord source connector, experimental Google Drive sign-in for a personal account, and experimental data dependencies for memoized functions
- ›The README badge says Python 3.10–3.13. The 1.2.0 package on PyPI (uploaded 2026-11-09) requires Python >=3.11, lists 3.11 through 3.14, and is classified Production/Stable. The separate cocoindex-code 0.2.43 package (uploaded 2026-11-05) is classified Alpha. Trust the package metadata over the badge
- ›The README says a source change updates the affected records and retires stale rows without touching anything that did not change. The core is Rust
Limitations
$ limits
- ›The README opens with '70% token savings' and does not link a benchmark or a method. Do not treat that line as a measured saving
- ›cocoindex-code[full] bundles sentence-transformers so embeddings run in-process, with no API key. The README says those dependencies are about 1 GB of torch and transformers. The slim install is LiteLLM-only: a cloud provider with an API key, or a local Ollama server with no key and no per-token cost. The local example is model: ollama/nomic-embed-text, and OLLAMA_API_BASE if the server is not at http://localhost:11434
- ›As an MCP server, ccc mcp exposes one tool, search. It does not look up callers or rename a symbol. The README also says ccc grep depends on a cocoindex code_match feature that, until it ships in a released cocoindex, requires a local cocoindex build
- ›With [full] installed, ccc init defaults to the in-process model Snowflake/snowflake-arctic-embed-xs. ccc init also recognizes the nomic-ai code models nomic-ai/CodeRankEmbed and nomic-ai/nomic-embed-code and fills in their indexing and query settings. The README shows those as sentence-transformers models, says CodeRankEmbed retrieves code better than the default, and says it is 137M parameters, needs about 1 GB of VRAM, and has an 8192-token context window
Sources
$ sources --verified
- overview2026-11-11cocoindex.io/
- audience2026-11-11cocoindex.io/
- features2026-11-11cocoindex.io/answers/cut-coding-agent-tokens/
- pricing2026-11-18cocoindex.io/enterprise/
- strengths2026-11-11github.com/cocoindex-io/cocoindex
- limits2026-11-11github.com/cocoindex-io/cocoindex-code
- why ranked2026-11-11cocoindex.io/answers/cut-coding-agent-tokens/
$ buildwith alternatives cocoindex
- #2 SerenaFree / OSS
use this if your coding agent needs to find a symbol, see who references it, and edit that symbol, instead of grepping and rewriting whole files.
- #3 SourcegraphPaid
use this if your team already has Sourcegraph, or will, and you want coding agents to search across repositories and get file ranges back instead of reading the tree themselves.
- #4 RepomixFree / OSS
use this if you want one file of a repository, or a compressed signature view of it, to hand to a chat model or an agent, and you do not want to run an index.