Repository-level documentation for large codebases, written by AI agents from the dependency graph.

Quick start • What's new in 2.0 • Features • Benchmark • Guides • Paper

📚 CodeWiki documents itself. Browse the documentation it generated for this repository at CodeWiki docs.

---

CodeWiki reads a repository, builds its dependency graph, splits the graph into a hierarchy of modules, and has agents write one page per module plus an overview, with Mermaid diagrams. It works on codebases from a few thousand to over a million lines, in 11 languages. It needs an LLM API key, or a Claude or Codex subscription, or an AI IDE that speaks MCP.

Quick start

1. Install

pip install git+https://github.com/FSoft-AI4Code/CodeWiki.git
codewiki --version

Needs Python 3.12+, Git, and Node.js with npm at install time.

2. Pick a provider

# Any OpenAI-compatible endpoint (OpenAI, LiteLLM proxy, OpenRouter, ...)
codewiki config set \
  --provider openai-compatible \
  --api-key YOUR_API_KEY \
  --base-url https://api.example.com/v1 \
  --main-model claude-sonnet-4 \
  --cluster-model claude-sonnet-4

# Or: no API key, run on your Claude Code subscription (`claude login` first)
codewiki config set --provider claude-code \
  --main-model claude-sonnet-4-6 --cluster-model claude-sonnet-4-6

Anthropic, Azure OpenAI, AWS Bedrock, Atlas Cloud, and Codex are also supported. See Providers and models.

3. Generate

cd /path/to/your/project
codewiki generate                     # writes ./docs/
codewiki generate --github-pages      # also writes an HTML viewer
codewiki generate --update            # later: refresh after code changes

What's new in 2.0

The documentation an agent can write is bounded by what the dependency graph contains. 2.0 makes that graph more complete, wider, and keeps it current. Full list in the CHANGELOG.

• More complete dependency graphs. Free functions become documentation units, so C code and function-centric C++, Java, and C# packages are no longer nearly empty. Call resolution in C, C++, Java, and C# is scope-, namespace-, include-, and import-aware. Library calls are filtered by an external symbol table. Ruby and Scala analyzers were added.
• Artifact-aware generation. Build files, CI workflows, Dockerfiles, package manifests, packaging scripts, configuration, and schemas become graph nodes with edges to the code they reference. They are clustered and documented like code, with a guaranteed Build, Deployment and Configuration module when clustering drops them. Guide
• Component-level incremental updates. codewiki generate --update diffs the saved graph against the current code, repairs the module tree, and sends one agent per affected module to patch its page and the pages that describe it. Everything else is left untouched. On svelte, one update after one commit cost $0.34 against $21.48 for a full build. Guide
• Subscription mode. Run on a Claude Pro/Max or Codex subscription through the claude and codex CLIs, with no API key.
• MCP server for IDE agents. codewiki mcp exposes the analysis toolchain to Cursor, Claude Desktop, Claude Code, or CodeBuddy. The IDE's own model does the writing. Guide
• Also: Atlas Cloud provider, prompt caching, .gitignore handling, Kotlin and PHP analyzers, a CI workflow, a security policy.

Features

Languages. Python, Java, JavaScript, TypeScript, C, C++, C#, Kotlin, PHP, Ruby, Scala.

Hierarchical decomposition. The dependency graph is clustered into a module tree, recursively, so a 1.4M-line repository gets the same treatment as a 10k-line one. Depth and size thresholds are configurable.

Recursive agents. One agent per leaf module reads the code through tools and writes the page. Parent pages are written from child pages. Diagrams are validated before they are saved.

Artifacts are documented. On by default. --no-artifacts gives the 1.x behaviour, --with-prose also reads README and docs/, --artifact-exclude skips generated configuration trees.

Incremental updates. --update refreshes only what changed. --compare-to <commit> sets the base commit for CI and squashed merges. If too much changed, it falls back to a full build and says so.

Customization.

codewiki generate --include "*.cs" --exclude "Tests,Specs,*.test.cs"
codewiki generate --focus "src/core,src/api" --doc-type architecture
codewiki generate --instructions "Focus on public APIs and include usage examples"
codewiki config agent --exclude "Tests,Specs"      # make it the default

--include replaces the default file set. --exclude merges with the built-in ignore list. Every flag is in the CLI reference.

IDE-driven mode. Add {"command": "codewiki", "args": ["mcp"]} to your IDE's MCP servers and ask the agent to document the repository. No LLM configuration in CodeWiki.

Output

./docs/
├── overview.md                  # start here
├── <module>.md ...              # one page per module, leaves and parents
├── module_tree.json             # the module hierarchy
├── first_module_tree.json       # clustering result before super-grouping
├── metadata.json                # model, version, commit, statistics
├── update_record.json           # every decision of the last --update run
├── temp/artifact_index.json     # artifact files by class
├── temp/dependency_graphs/      # the saved graph (used by --update)
└── index.html                   # viewer (with --github-pages)

This repository's own output is checked in under ./docs/.

Benchmark results

Evaluated on CodeWikiBench: seven repositories, 486 rubric requirements, the same LLM judge for every system. Scores are weighted rubric scores from 0 to 100. DeepWiki and CodeWiki 1.0 were re-run under the same setting as 2.0 (September 2026), so the numbers differ from the ones in the paper.

Repository	Language	LoC	DeepWiki	CodeWiki 1.0	CodeWiki 2.0	2.0 vs 1.0
OpenHands	Python	229,909	74.12	83.31	83.53	+0.22
svelte	JavaScript	124,576	69.83	73.21	80.09	+6.88
puppeteer	TypeScript	136,302	65.57	84.16	85.04	+0.88
ml-agents	C#	86,106	75.74	80.92	89.35	+8.43
logstash	Java	117,485	56.31	59.42	77.85	+18.43
Wazuh	C	1,446,730	69.91	65.38	88.14	+22.76
Electron	C++	184,234	43.92	42.12	71.59	+29.47
Average			65.06	69.79	82.23	+12.44

The gain is largest where the 1.0 graph missed the most structure (C, C++, Java, C#) and near zero where the analyzers did not change (Python, JavaScript, TypeScript). Artifacts lift the build and testing topics, which 1.0 could not see at all.

Topic	Requirements	DeepWiki	CodeWiki 1.0	CodeWiki 2.0
Architecture	128	72.95	78.02	82.00
Functionality	211	61.62	70.46	88.65
Operations	63	67.71	78.24	78.10
Build and deployment	27	62.84	28.16	71.90
Security	29	48.73	64.91	83.20
Testing	12	61.43	41.93	63.25
Performance	16	43.33	49.06	70.90

(a) how much the 2.0 code graph grew over 1.0; (b) nodes and edges added by artifacts; (c) score by stage. "Conference version" is CodeWiki 1.0, "C1 + C2" is CodeWiki 2.0.

Requirements

• Python 3.12+
• Node.js and npm at install time (a dependency builds against them)
• Git
• One of: an LLM API key, a Claude Code or Codex subscription, or an MCP-capable AI IDE

Guides

CLI reference	every command and flag
Providers and models	API keys, Atlas Cloud, Azure, Bedrock, subscription mode
Artifact-aware generation	what gets documented beyond code, and how to tune it
Incremental updates	how --update works, thresholds, the update record
MCP / IDE-driven mode	Cursor, Claude Desktop, Claude Code, CodeBuddy
Development guide	layout, pipeline, adding a language, tests, releasing
Docker setup	the web application in a container
CodeWikiBench	the benchmark
Live demo	generated documentation examples

Citation

CodeWiki was introduced in CodeWiki: Evaluating AI's Ability to Generate Holistic Documentation for Large-Scale Codebases, published in Findings of ACL 2026. If you use CodeWiki in your research, please cite:

@inproceedings{hoang-etal-2026-codewiki,
    title = "{C}ode{W}iki: Evaluating {AI}{'}s Ability to Generate Holistic Documentation for Large-Scale Codebases",
    author = "Hoang, Anh Nguyen and Le-Anh, Minh and Le, Bach and Bui, Nghi D. Q.",
    booktitle = "Findings of the {A}ssociation for {C}omputational {L}inguistics: {ACL} 2026",
    month = jul,
    year = "2026",
    address = "San Diego, California, United States",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2026.findings-acl.288/",
    doi = "10.18653/v1/2026.findings-acl.288",
    pages = "5812--5827",
}

Star history

Sponsors

CodeWiki is proudly sponsored by FPT Software.

License

MIT. See LICENSE.