CodeNib¶

CodeNib finds the source code a coding agent needs and serves it with file and line evidence. It builds several views of one repository snapshot (BM25, embeddings, a typed SCIP/LSP symbol graph) and offers them three ways: as fixed retrieval pipelines, as fixed LLM workflows, and as tools inside an agent loop.

The shipping path for Codex and Claude Code is local CodeGraph setup: model-free, with no API key. The headline measured result is the grep → Jev experiment: 71.4% code-block Recall@5 on 100 issues, 12.8 percentage points above the same grep candidates without reranking. That route is optional and uses remote models.

How these docs are organized¶

Section	Read it to	Start with
Get Started	Install CodeNib and connect a repository to your agent	Agent-ready CodeGraph
Guides	Serve context over MCP, add grep → Jev, or embed CodeNib in an agent	MCP Server
Concepts	Understand the views and how each method uses them	Context methods and support matrix
Benchmarks	See what is measured, on which datasets, and how to reproduce it	Method × benchmark matrix
API Reference	Browse the public Python contracts	`codenib` public surface
Development	Extend, test, and release CodeNib	Contributing a Language

Language support differs by view. The generated language capability matrix shows which languages support chunking, graph indexing, incremental backends, and C++ core decoder parity.

Quick Start¶

CodeNib 0.2.4 can prepare a local, model-free CodeGraph and register it with installed Codex and Claude Code clients in one command:

python -m pip install "codenib[graph,mcp]==0.2.4"
codenib codegraph init /path/to/repository

The command installs package-managed language providers, builds reusable BM25 and symbol-graph views under ~/.codenib, and delegates MCP registration to the clients' own CLIs. The target checkout remains unchanged. Read the CodeGraph guide for status, safe uninstall, client scopes, and tool examples.

python -m pip install "codenib[semantic]==0.2.4"
codenib wiki /path/to/repository

The separate Wiki path enables hybrid BM25+dense retrieval with the pinned CodeRankEmbed model. Read the Quickstart for provider setup and browser troubleshooting.

Serving Docs Locally¶

git clone https://github.com/sysevol-ai/CodeNib.git
cd CodeNib
python -m pip install -e ".[dev]"
mkdocs serve

Then open http://localhost:8000.