CodeNib

CodeNib is a multi-view data system for serving repository context to coding agents. It compiles repositories into reusable, source-linked lexical, semantic, structural, and static-navigation views, then serves the same manifest through MCP, Python APIs, the local Wiki, and evaluation harnesses.

Language support varies by surface. Start with the generated Language Capabilities matrix when you need to know which languages support chunking, graph indexing, incremental patching, or C++ core decoder parity.

User Guide

Goal	Read
Turn a local repository into a Wiki	Quickstart
Build an index and expose it to an agent	MCP Server
Configure the full browser application	Web UI
Run without cloud-hosted LLM APIs	Running Locally
Understand language support and gaps	Language Capabilities
Add or promote a language backend	Contributing a Language

Once an index exists, use Agent Skills for optional retrieval and reranking workflows, or RAG Ops And Planner for the query planner and graph-aware retrieval boundaries.

Concepts

Surface	Description
Index and MCP	Build manifests with IndexCompiler; serve BM25, semantic, regex, Zoekt, and dependency-subgraph tools over MCP
Web UI	Source-linked Wiki and optional Ask flow over indexed repositories
Graph Range Query	LSP-aligned line-range and symbol queries with typed, source-linked results
Incremental Graph	Update a graph in place after a diff where the language backend supports it
SCIP Index	SCIP and language-server graph indexing details, cache levels, and backend behavior
Core C++ Backend	libigraph-based decoder acceleration and parity boundaries

Development

• Contributing a Language — add a language through the shared registry and verify each supported surface.
• CI/CD — run local and remote test tiers, including serial, graph-consumer, core, and slow jobs.
• Releasing — prepare and verify a CodeNib release.
• Naming — use the project and package names consistently.

Evaluation

• Collecting SWE-bench Instances — sample representative instances across languages.
• Synthesis Pipeline — synthesize traversal, behavior, and multi-step queries for the CodeNib Synthesis benchmark.
• GT Locator — evaluate source-location retrieval against ground-truth patches.
• Evaluation Artifact Bundles — package reproducible evaluation outputs.

Quick Start

pip install codenib
codenib wiki /path/to/repository

The default Wiki is deterministic, uses BM25, and needs no API key. CodeNib detects repository languages, writes a reusable manifest under ~/.codenib/repositories, and opens the browser UI. This default fast profile leaves the target checkout unchanged. Read the Quickstart for profiles and troubleshooting.

pip install "codenib[mcp]"
codenib mcp /path/to/repository

Serving Docs Locally

pip install mkdocs-material
mkdocs serve

Then open http://localhost:8000.