CodeNib README

The full upstream README, mirrored for reference. Install config, tool schemas, adoption signals and our own overview live on the CodeNib listing page.

CodeNib

Understand unfamiliar code with evidence.

Read a Wiki of how a repository works, follow its call paths on a map, and open the source behind every explanation.
Then give your coding agent the same index.

Browse a Wiki  ·  Use with your agent  ·  Thread context  ·  Docs  ·  Languages  ·  Releases

CI PyPI version Latest GitHub Release arXiv paper Hugging Face: #1 Paper of the Day Apache 2.0 license

Claude Code plugin Codex plugin Grok Build plugin DeepSeek Harness plugin CodeNib Wiki action on GitHub Marketplace

The CodeNib Wiki for psf/requests: the page list on the left, a map of the Wiki's areas joined by recorded references in the middle, and the Redirect Resolution page on the right, which traces Session.send() to resolve_redirects() to get_redirect_target() with a file and line for each step

The Requests Wiki. Each node on the map is a page, and the heavy lines are references the code index recorded between areas. The page cites a file and line range for each claim.

New in 0.2.6: Wikis open as a map of their pages and lead with the main path through the code, one GitHub Action publishes yours to GitHub Pages, plugins connect Claude Code, Grok Build and DeepSeek Harness, graph builds get faster, and first C/C++ and JavaScript/TypeScript builds leave the checkout alone. Two previews serve coding agents: codenib guard lists the tests a change can reach, and codenib threads hands your agent a GitHub issue or pull request in one read. Read the release notes.

A Wiki you can follow down to the source

CodeNib indexes a repository, groups its code into areas, and writes a Wiki page for each one. The map and the call paths come from the code index, so you can check every arrow against the source.

  • Start from the map. Pages are nodes. Drag to pan, scroll to zoom, and click a node to read its page. Click a heavy line to read an example of the references behind it.
  • Follow one path through the code. The Overview opens on the main path through the repository, one sentence per recorded call. The most important detail pages draw their longest recorded call path. No model chooses the arrows.
  • Check each claim. Sentences end with the file and line range they come from, and each page lists its callers and calls in the code graph.
  • Ask a follow-up. Ask this repository answers a question with source citations, using the model you configure.

The CodeNib Wiki for gin-gonic/gin: a map of eight areas such as Radix Routing Tree and Engine Lifecycle, beside the Overview, which follows one recorded path from GET() through RouterGroup.handle and Engine.addRoute() to node.addRoute, citing the file and line of each call

The gin Wiki opens on one recorded path, from GET() to node.insertChild().

Open a Wiki

ToDo this
Read one nowOpen a ready Wiki for requests, scikit-learn, sympy, fmt, redis, caddy, gin, hugo and more. No account or API key.
Map a public GitHub repositoryPaste its URL on demo.codenib.ai, or replace github.com with demo.codenib.ai in the URL. A new Wiki uses your OpenRouter account, with a $0.25–$5 spending limit you set before it starts.
Map your own checkoutRun the Wiki locally, as below.
Share itPublish the Wiki, with its source citations, to GitHub Pages: add the CodeNib Wiki action to a workflow. Pages guide

To build a Wiki with the map on your machine:

bash
python -m pip install "codenib[graph]"
codenib wiki /path/to/your/repo --preset graph

This builds a search index and a symbol graph, then opens the Wiki at http://localhost:3000. Pages are written from the index without a model. Add --generate --model <provider/model> for model-written pages. Local Wiki guide

Use it with your coding agent

Build the same index for Claude Code or Codex. Your agent can then find definitions, follow callers and read the source in your own checkout.

The callers of rebuild_auth() in Requests, as one dependency_subgraph call returned them. Eight public entry points, api.request() and Session.get(), options(), head(), post(), put(), patch() and delete(), each reach Session.request() at sessions.py:500. From there one path leads through Session.send() at sessions.py:673 and resolve_redirects() at sessions.py:159 to rebuild_auth() at sessions.py:282.

One recorded Claude Code session on Requests. Asked where Requests strips authentication on a redirect to another host, and which public methods reach that code, the agent found rebuild_auth() with explore_context. One dependency_subgraph call then returned the path above, each step with its file and line. The session took 14.1 s, 3 turns and 2 CodeNib tool calls. ▶ Watch it replay · Pinned source, transcript and setup details

Requires Python 3.10+, Git, and Claude Code or Codex. Initialization accepts uncommitted edits and preserves the checkout.

bash
python -m pip install "codenib[graph,mcp]"
codenib codegraph init /path/to/your/repo

Then ask your agent: “Use CodeNib's explore_context to find where request retry behavior is implemented. Cite the files and lines.”

This builds local search and a typed symbol graph, then connects installed agent clients. CodeNib needs no model, API key, or GPU of its own; your agent uses its own model. Language toolchain and project prerequisites vary.

In Claude Code you can install the plugin instead. It connects every Git checkout you open and adds /codenib:explain, /codenib:impact, /codenib:thread, /codenib:setup and /codenib:wiki:

Terminal
claude plugin marketplace add sysevol-ai/CodeNib
claude plugin install codenib@codenib

Codex and Grok Build install the same plugin, and DeepSeek Harness installs it as a bundle:

bash
codex plugin marketplace add sysevol-ai/CodeNib --sparse .claude-plugin --sparse plugins/codenib
codex plugin add codenib@codenib
grok plugin install "sysevol-ai/CodeNib#plugins/codenib"
dsh plugin --profile web add "git+https://github.com/sysevol-ai/CodeNib.git#main&path:/plugins/dsh"

Codex setup · Grok Build setup · DeepSeek Harness setup

Setup and troubleshooting · Source exclusions · All MCP tools · Change guard (preview, Python: lists the tests a change reaches before your agent finishes)

Hand your agent the whole thread

codenib threads reads a GitHub issue or pull request once and writes one document for your agent: the ask, the review comments, the failing checks, the linked threads, and the code they point at. It runs on your machine and calls no model. It is a preview in 0.2.6.

Change in input tokens when the agent is handed the thread context, against reading the thread itself with gh. Codex on 76 pull requests: 21% fewer, with a 95% interval of 27% to 14% fewer. Codex on 47 issues: 25% fewer, 31% to 19%. Claude Code on 76 pull requests: 45% fewer, 57% to 30%.

Each dot is the change in input tokens over all the tasks when the agent is handed the prepared context, against reading the thread itself with gh, and each line is its 95% interval. An agent that reads a thread itself makes one call per piece, and every call sends the conversation so far again; the prepared document is read once. The agents found the files to change about as often either way: 0.75 against 0.74 and 0.85 against 0.82 with Codex, 0.80 against 0.83 with Claude Code.

bash
cd /path/to/your/repo
codenib threads context 812     # write THREAD_CONTEXT.md before the agent starts
codenib threads open 812        # or start a ticket that remembers what it read
codenib threads inbox 812       # later: only what is new
codenib threads board           # all tickets, and other work on the same code

When several agents work on one repository, each ticket is sent only the records it has not read, and tickets whose work reaches the same function are told about each other.

Thread context guide · How it decides what to include, with real output · In Claude Code and Grok Build, the plugin adds it as /codenib:thread.

How it compares

What an agent can ask each tool. Peer interfaces were checked against upstream docs on 2026-09-28. This compares capabilities, not benchmark scores.

ToolAsk in plain languageRelationships beyond one hopRelationships resolved byRuns
grep / readNo; literal or regexNoText matchLocal
CodeNib CodeGraphYes; ranked source anchors (explore_context)Incoming/outgoing static references; one dependency_subgraph call, depth ≤ 8SCIP/LSP indexers (compiler-resolved)Local
SerenaNo; symbol name or regexOne level per call (find_referencing_symbols)Live language server or IDELocal
CodeGraph (colbymchenry)Yes; FTS5 full-text (codegraph_explore)Yes; call paths in codegraph_exploreTree-sitter AST extractionLocal; telemetry opt-out
DeepWiki public MCPYes; generated answers (ask_question)No structured graph toolsGenerated WikiHosted

CodeNib's indexed references can include calls, fields, type uses and function values. Inspect the source to establish which references invoke a function; external and dynamic relationships may be missing. For a known string or file, your agent's native search and reads remain useful. Include indexing and rebuild costs when assessing graph navigation on your own tasks.

None of these repository tools needs its own model or API key; your agent uses its own model. Detailed comparison, sources and boundaries. CodeNib also includes an optional grep → Jev route using OpenRouter for planning and reranking; selected code goes to remote models. Authorization previews remain opt-in.

Languages

CodeNib registers 14 languages, 12 of them with a code graph.

SupportLanguages
Code graph with an incremental backend (LSP or clangd)Python, Go, Rust, C/C++, JavaScript, TypeScript
Code graphC#, Java, Ruby, PHP, Kotlin, Scala
Search only, no definitions or callersSwift, Lua

These are registered capabilities, not the tools on your machine; codenib toolchain status shows those. The product currently reuses or rebuilds views; file-level delta repair is not enabled. The language matrix lists each backend and its prerequisites.

Results and reproductions

  • #1 in line recall and Pass@1 on ContextBench's Agent systems leaderboard: 77.90% line recall and 388 of 500 tasks resolved (77.60% Pass@1), as of October 23, 2026 (UTC). This is our research agent, not the Wiki or the code graph: two Claude Haiku 5.5 workers search with glob, grep and reads, then Claude Opus 5.5 repairs. How it works · Scope and evidence
  • 71.4% code-block Recall@5 on 100 real issues with model-planned grep → Jev, 12.8 percentage points above the same grep candidates without reranking. This is an experimental retrieval route, separate from the model-free CodeGraph. Method and limitations
Start hereWhat you can inspect
Context methods and support matrixEvery retrieval pipeline, LLM workflow and agent route, sorted by the views it reads and who controls the search
BenchmarksLocalization, agent-context and resolution layers, evidence levels, and the method × benchmark matrix over CodeNib Base/Synthesis, SWE-bench Lite, Loc-Bench, SWE-Explore, Agent Retrieval Bench and ContextBench
Agent integration matrixRevision-pinned LocAgent, Agentless, CoSIL, OrcaLoca and RepoNavigator contracts; compatibility does not imply reproduced paper scores
SWE-Explore validation1,020/1,020 real-output metric cells match the pinned official evaluator on a fixed 20-case run
DGX Spark deploymentLocal Wiki, CodeGraph and model serving on GB10; a deployment guide, not a token-saving benchmark

Documentation and community

Documentation · Architecture · Contributing · CI and testing · Changelog · Discord

CodeNib is in beta; public interfaces may change before a stable release. Historical research artifacts retain their published dataset identifiers.

Citation

If you use CodeNib in your research, please cite our arXiv paper:

bibtex
@misc{yu2026codenibmultiviewdataserving,
      title={CodeNib: A Multi-View Data System for Serving Repository Context to Coding Agents},
      author={Zhongming Yu and Hengjia Yu and Boqin Yuan and Shuting Zhao and Yizhao Chen and Aryan Dokania and Mihir Jagtap and Jiayu Chang and Yitong Ma and Yash Jayswal and Wentao Ni and Hejia Zhang and Zhaoling Chen and Gangda Deng and Jishen Zhao},
      year={2026},
      eprint={2607.25431},
      archivePrefix={arXiv},
      primaryClass={cs.SE},
      url={https://arxiv.org/abs/2607.25431},
}

CodeNib is licensed under Apache 2.0.