AGPL-3.0 · Open Source · Self-hostable
Your codebase,
finally understood.
repowise parses your code into a dependency graph, mines git history for ownership and churn, generates a wiki with an LLM, and exposes everything through 8 MCP tools your AI agent can query in real time.
One engine, three interfaces
Install once. Choose the interface that fits your workflow — or use all three. They share the same data, the same intelligence, the same stores.
CLI
For the solo developer
pip install repowise. Run init, update, search, dead-code, and 10 more commands. Works fully offline with Ollama. Your code never leaves your machine.
MCP Server
For AI-native workflows
8 tools that plug into Claude Code, Cursor, or Cline. Your AI agent calls get_context() instead of reading 40 files. Config auto-generated after repowise init.
Web UI
For the whole team
Browse the wiki, explore the dependency graph in D3, view hotspot tables, track doc freshness, and chat with your codebase. All served from repowise serve.
From install to codebase intelligence
in four commands
One pip install. Works offline with Ollama. No account required to get started — run everything on your machine.
Install
Works on Python 3.10+. Supports Anthropic, OpenAI, Gemini, and Ollama (fully offline).
Index your repo
Parses your codebase, builds a dependency graph, mines git history, and generates wiki pages with an LLM.
Stay current
Run after commits. Detects changed files, cascades updates through the dependency graph.
Open the web UI
Full dashboard: wiki explorer, interactive dependency graph, hotspot tables, ownership maps, and codebase chat.
Most tools answer one question.
repowise answers four.
Graph structure, git history, generated documentation, and architectural decisions — four layers that compound into genuine codebase understanding.
Every dependency, ranked and traced
repowise parses your codebase into a directed dependency graph using tree-sitter ASTs across 10 languages. PageRank identifies your most critical symbols. Community detection discovers logical modules even when directory structure doesn't reflect them.
- 6,284+ nodes with PageRank, betweenness centrality, SCC detection
- Edge types: imports, calls, inherits, implements, co-changes
- Topological sort drives generation order — bottom-up, always
- Scales to 30K+ nodes with automatic SQLite-backed graph
History that writes the documentation
GitIndexer mines your commit history to classify files as hotspots or stable, compute ownership, extract significant commit messages, and discover co-change partners. These signals flow into generation prompts, so your wiki explains why code was written — not just what it does.
- Hotspot detection: top 25% churn + complexity files flagged
- Co-change partners: files that change together without imports
- Ownership from git blame — primary owner + top 3 contributors
- Significant commits filtered and included in generation prompts
Wiki pages that stay fresh
Each wiki page is generated with 9 layers of context: source code, symbol signatures, graph metrics, git history, import summaries, RAG context, co-change docs, dead code findings, and reverse imports. Confidence scores decay when source changes — stale pages auto-regenerate.
- Confidence scoring: 0.0–1.0 with git-informed decay modifiers
- RAG context via LanceDB or pgvector — each page knows its imports
- 9-level hierarchical generation: symbols → files → modules → repo
- Resumable jobs — crash-safe, idempotent, checkpoint after every page
The why behind your architecture
Decisions are extracted from four sources: inline markers (# WHY:, # DECISION:, # TRADEOFF:), git archaeology, README/docs mining, and manual CLI capture. Each decision tracks staleness — when affected files change, the decision is flagged for review.
- 4 capture sources with confidence: inline (0.95), git (0.70–0.85), docs (0.60), CLI (1.00)
- Staleness tracking — decisions age when governed files get commits
- Health dashboard: stale decisions, ungoverned hotspots, proposed reviews
- MCP tool get_why() searches decisions before you change anything
CLAUDE.md that writes itself
After every repowise update, the CLAUDE.md is regenerated from real data — graph metrics, git history, dead code findings. No LLM needed. No templates. Just your codebase, distilled into the context your AI agent needs.
- Architecture overview extracted from the real dependency graph
- Hotspot warnings with file paths, churn metrics, and owners
- Architectural constraints and key design decisions
- Dead code summary with confidence scores
- Entry points, build commands, and tech stack detection
Also generates cursor.md for Cursor users. Same data, different format.
8 tools your AI agent already knows how to call
The MCP server exposes the entire wiki as structured, queryable tools. Instead of reading 40 source files, your agent calls get_overview() and gets a current architecture summary. Config auto-generated for Claude Code, Cursor, and Cline.
get_overview()Architecture summary, module map, entry points, tech stack.
First call when exploring an unfamiliar codebase.
get_context()Docs, ownership, history, decisions, freshness for files, modules, or symbols. Pass multiple targets in one call.
Before reading or modifying specific code.
get_risk()Hotspot score, dependents, co-change partners, risk summary. Also returns top 5 global hotspots.
Before modifying files — assess blast radius.
get_why()Three modes: natural language search over decisions, path-based lookup, or health dashboard.
Before architectural changes — understand existing intent.
search_codebase()Semantic search over the full wiki using LanceDB or pgvector. Natural language queries.
When you don't know where something lives.
get_dependency_path()Connection path between two files or modules in the dependency graph.
Understand how two things are connected.
get_dead_code()Unreachable files, unused exports, zombie packages — sorted by confidence and cleanup impact.
Before cleanup or refactoring tasks.
get_architecture_diagram()Mermaid diagram for the full repo or a specific module scope.
For documentation, presentations, or onboarding.
What your git history has been trying to tell you
repowise mines commit history into actionable signals: which files break most often, who owns what, and which files are secretly coupled. These signals drive generation depth, confidence decay, and risk assessment.
Hotspot Detection
Top 25% churn + complexity files flagged automatically. Hotspot pages get deeper generation and faster confidence decay.
Ownership Maps
git blame + commit history reveals who owns what. Bus factor calculated per module. Single-owner modules highlighted.
Co-change Partners
Files that change together without import relationships. These hidden couplings are invisible to static analysis.
Significant Commits
Large diffs, refactors, and migration commits are identified and included in generation prompts for richer context.
Confidence Decay
When source files get new commits, associated wiki pages decay in confidence. Stale pages auto-regenerate on next run.
90-Day Windows
All git metrics use rolling 90-day windows by default. Reflects current activity, not ancient history.
Ask questions.
Get grounded answers.
The chat agent has access to all 8 MCP tools. It searches the wiki, reads the graph, checks git history, and cites architectural decisions — all automatically. Answers are grounded in your actual codebase, not hallucinated from training data.
- Provider-agnostic — uses whichever LLM you configured
- SSE streaming with real-time tool call visibility
- Conversation persistence across page refreshes
- Artifact panel for rich results: wiki pages, diagrams, risk reports
Team intelligence, managed for you
Everything in the open-source product, plus team features. No infrastructure to manage. We handle ingestion, storage, and updates.
Shared Team Context
Everyone on the team queries the same wiki, same graph, same decisions. New hires get onboarded with get_overview() instead of a stale Notion doc.
Plugin Store
COMING SOONExtend repowise with community plugins: custom parsers, integration hooks, specialized MCP tools, and export formats.
Role-Based Access
COMING SOONControl who can trigger regeneration, view sensitive repos, or modify architectural decisions. SSO integration included.
Analytics Dashboard
COMING SOONTrack doc freshness trends, generation costs, token usage, hotspot evolution over time, and team adoption metrics.
The full picture, side by side
Most tools solve one slice of the problem. repowise is the only open-source platform that combines auto-generated documentation, git intelligence, decision records, and MCP tools in a single self-hostable package.
| Feature | repowise | Google CodeWiki | DeepWiki | CodeScene | Sourcegraph |
|---|---|---|---|---|---|
| Self-hostable OSS | ✓ | — | — | — | — |
| Works with private repos | ✓ | — | ✓ | ✓ | ✓ |
| Auto-generated wiki (LLM) | ✓ | ✓ | ✓ | — | — |
| Git intelligence (hotspots / ownership / co-changes) | ✓ | — | — | ✓ | — |
| Dead code detection | ✓ | — | — | — | — |
| Architectural decision records | ✓ | — | — | — | — |
| MCP server for AI agents | ✓ | — | — | — | — |
| Semantic search | ✓ | ✓ | ✓ | — | ✓ |
| Doc freshness / confidence scoring | ✓ | — | — | — | — |
| CLAUDE.md auto-generation | ✓ | — | — | — | — |
| Codebase chat (agentic) | ✓ | ✓ | ✓ | — | — |
| Dependency graph visualization | ✓ | ✓ | ✓ | ✓ | ✓ |
| Provider choice (4 LLM providers) | ✓ | — | — | — | — |
| Privacy (code never leaves your infra) | ✓ | — | — | ✓ | ✓ |
Guides, comparisons, and deep dives
Learn about codebase intelligence, MCP tools for AI agents, and how repowise compares to other tools.
5 Best Codebase Documentation Tools in 2026 (Compared)
For decades, 'good documentation' was the white whale of software engineering. We all knew it was necessary, but the friction of maintaining it—manually upda...
repowise vs CodeScene: Full Codebase Intelligence Without the Enterprise Price Tag
Technical debt is rarely a single catastrophic event. Instead, it’s a slow accumulation of 'paper cuts'—a complex module here, a high-churn file there, and a...
repowise vs DeepWiki: Self-Hosted Codebase Docs That Stay Fresh
Every engineering team eventually hits the 'Documentation Wall.' It’s the moment when the codebase grows faster than the human capacity to describe it. You’v...
Three paths to codebase intelligence
- Self-host (free, forever)
pip install repowise. Run on your machine, your server, your CI. AGPL-3.0. Full feature set. Your code never leaves your infrastructure. - Hosted (coming soon)
Managed infrastructure, team features, shared context. Join the waitlist and we'll notify you when it's ready.
- Enterprise
On-prem deployment, SSO, role-based access, dedicated support, SLAs. Reach out and we'll scope it together.