Who repowise is for#
AI coding teams using Claude Code, Cursor, or Codex
Teams whose agents guess at which files matter, change callers they never found, and generate diffs requiring lengthy reviews. repowise gives agents a pre-built dependency graph, decision records, and health scores via ten MCP tools, reducing output tokens 31.6% and tool calls from 7.2 to 3.8 in a measured benchmark.
Skip if:
Teams not yet using AI coding agents or MCP-compatible editors. The MCP tools are the primary value delivery mechanism; without an MCP client, only the CLI and web UI are available.
Solo developers auditing a complex or inherited codebase
A developer taking over an unfamiliar repo gets a dependency graph, file-level health scores with concrete refactoring plans (Extract Class, Break Cycle, Split File), ownership history, and a generated wiki in one pip install. The first index needs no API key.
Skip if:
Developers working on small or greenfield projects where the architecture is already well-understood. repowise's value grows with codebase complexity and age.
Platform engineers tracking ownership and bus factor
repowise computes ownership percentages and bus factor from full git blame across the org's codebase and surfaces co-change pairs that reveal hidden coupling. Portfolio health dashboards show which teams own the highest-risk files. The Teams plan at $20/seat/month covers a shared index and credit pool.
Skip if:
Teams on a single small repository with a stable, well-documented ownership structure.
Enterprise teams requiring air-gapped deployment
The enterprise tier supports on-premises deployment inside a private network with zero telemetry, no phone-home, BYO LLM or fully offline mode via Ollama. A commercial license removes the AGPL obligation and adds IP indemnification, a defensive patent grant, SAML/OIDC SSO, and SCIM provisioning.
Skip if:
Teams comfortable with AGPL-3.0 terms and a cloud-connected deployment. The enterprise tier addresses compliance, air-gap, or IP requirements; the self-hosted AGPL tier covers all other cases.
The problem it solves#
AI coding agents make teams fast at first, then slower. Agents grep through codebases to find relevant files, change functions with dozens of callers they never discovered, and generate diffs that senior engineers spend hours reviewing because nobody has a shared model of what the code does or why it is shaped that way.
These are knowledge problems, not model problems. A repository's dependency graph, change history, architectural decisions, and code health signals are all computable ahead of time. Instead, agents rediscover them from scratch on every task: grep, read, re-read, forget. The exploration phase burns tokens and tool calls on work the next agent will repeat in full. Review becomes the bottleneck because generation got fast but confidence did not.
How it solves it#
Ten MCP tools for agent context
Ten task-shaped MCP tools expose the codebase index to Claude Code, Cursor, Codex, and any MCP-compatible client. Tools are built around tasks rather than data entities: pass several targets in one call and get the complete context, dependency graph, health, and architectural decisions back in one response rather than chains of sequential file reads.
Defect-validated code health scoring
Every file is scored 1-10 by 49 deterministic detectors across three lenses: defect risk, maintainability, and static performance risk. The score is validated at ROC AUC 0.74 across 21 repositories and 9 languages, with repowise-flagged files containing 2.3x the defects of a leading commercial tool under the same review budget. No LLM calls; runs in under 30 seconds on a 3,000-file repo.
Architectural decision records
Decisions are mined from five sources: pull requests, inline markers, README archaeology, CLI captures, and documentation generation. Each decision is evidence-backed, traced to a verbatim source span, and attached to the files it governs. MCP tools deliver the decision to an agent or reviewer before a change is made rather than after.
Dependency graph across 18 languages
Tree-sitter ASTs across 18 languages resolve into a directed graph of imports, calls, inheritance, and co-changes. PageRank and betweenness centrality rank structurally load-bearing files, so blast radius from a planned change is computed rather than guessed. Route-to-handler edges span 22 frameworks.
Git intelligence: hotspots, ownership, and co-change pairs
Hotspots combine decayed churn against complexity to surface files accumulating risk. Ownership percentages and bus factor are computed from full git blame. Co-change pairs surface hidden coupling: files that move together across commits with no import between them. All signals filter out doc, test, and config commits so the counts mean what they say.
Command output distillation
The distill command compresses command output before an agent reads it, keeping errors first and exit code intact. On a pytest run it delivers 61% fewer tokens; on git log it reaches 89% fewer tokens. Every omission leaves an inline reference marker that the agent can expand on demand, so no information is lost.
Strengths and trade-offs#
Strengths
- Zero LLM calls in the indexing pipelineEvery intelligence layer (graph, health, git, documentation, decisions) is computed deterministically over tree-sitter ASTs and git data. No model call runs during indexing, which means results are reproducible, auditable, and free from model hallucination in the analysis path. EU AI Act high-risk obligations do not apply because scoring is fully deterministic.
- Benchmarks published with losses includedrepowise publishes its benchmark harness, corpus, and sample sizes, including the rows where a competitor wins. In file-retrieval accuracy it ranked #1 of 6 at 0.876 file coverage against the next tool's 0.610 on a sealed 42-instance split (p=0.00004). CodeGraph is a genuine second at -24.4% token reduction, and that result is reported rather than buried.
- Self-hosted with zero telemetrySource code is processed transiently and never persisted outside the graph and git metadata. Self-hosting requires no cloud account, no API key for the first index, and no external calls. Enterprise deployment supports fully air-gapped networks with BYO LLM or Ollama for fully offline operation.
- CLAUDE.md generated from the real indexrepowise generates your CLAUDE.md and AGENTS.md from the live codebase index rather than templates, so agents without MCP support still start each session with accurate architecture context. The wiki generation budget tilts toward the modules you and your agents ask about most, based on actual usage patterns read from your own agent transcripts.
Trade-offs
- -AGPL-3.0 requires source disclosure for network servicesThe core engine is AGPL-3.0. Teams that modify and run repowise as a service for others must publish their source changes under the same license. Commercial users who need proprietary modification rights, embedding, or IP indemnification require the enterprise commercial license.
- -Indexing speed is the slowest in its benchmark fieldrepowise publishes this itself: it is the slowest indexer in its benchmark comparison. The under-30-second figure applies to incremental updates on a 3,000-file repo, enforced by a CI test. Cold initial indexing time scales with repository size and may be a consideration for very large monorepos.
- -104 open GitHub issues at time of indexingWith 5,441 stars and rapid adoption since its March 2026 launch, the issue tracker has 104 open items. This is expected for a tool gaining traction quickly. Users building workflows that depend on specific language support or edge-case handling should review open issues for anything relevant to their setup before committing.
repowise vs alternatives#
repowise vs CodeScene
CodeScene is a proprietary code health and analysis platform used by enterprise engineering teams. Both tools score code health and surface technical debt. The core difference is deployment model and scope.
| Feature | repowise | CodeScene |
|---|---|---|
| License | AGPL-3.0 (self-hosted) | Proprietary |
| Self-hosting | Yes | Limited |
| MCP server for AI agents | Yes (10 tools) | No |
| Architectural decision records | Yes | No |
| Defect-validated scoring | ROC AUC 0.74, published | Not published |
| Git analytics | Yes | Yes |
repowise's code health score uses 49 deterministic detectors with no LLM calls, validated at ROC AUC 0.74 across 21 repos and 9 languages. In a head-to-head on the same 2,770 files under the same review budget, repowise surfaced 2.3x more defects. CodeScene is still the stronger choice for organizations already embedded in its workflow, teams that need its specific process-metric visualizations, or buyers who prefer a fully managed SaaS with no infrastructure to operate.
repowise vs DeepWiki
DeepWiki generates wiki-style documentation for GitHub repositories as a hosted service. Both tools produce auto-generated documentation. The difference is deployment model, depth, and breadth.
| Feature | repowise | DeepWiki |
|---|---|---|
| License | AGPL-3.0 (self-hosted) | Proprietary (hosted only) |
| Self-hosting | Yes | No |
| MCP tools | Yes (10) | No |
| Decision records | Yes | No |
| Code health scoring | Yes (defect-validated) | No |
| Doc freshness scoring | Yes | No |
| Free tier | Yes (self-hosted) | Yes (public repos) |
repowise generates documentation bottom-up from symbols to modules, rebuilds incrementally on every commit, and assigns freshness and confidence scores to every page. DeepWiki is faster to start: paste a GitHub URL and the wiki is ready with no install required. DeepWiki is the better choice for quickly exploring an unfamiliar public repository without any setup. repowise is the better choice when documentation must stay current automatically, serve the same index to AI agents via MCP, or combine with code health and decision records in one self-hosted deployment.
Install and self-host#
Install repowise via pip, then run init to index your repository and serve the local dashboard.
```bash
pip install repowise
repowise init .
repowise update
repowise serve
```What it's built on#
- Languages
- PythonTypeScript
- Frameworks
- React
FAQ#
Does source code leave my machine when using repowise?
No. repowise is self-hosted with zero telemetry, and source code is processed transiently and never persisted. What is stored is the dependency graph, non-reversible embeddings, generated wiki pages, and git metadata. You can bring your own LLM API key or run fully offline using a local model via Ollama.
How much does repowise reduce AI agent token usage?
Oops! Something went wrong
[next-mdx-remote-client] error compiling MDX: Unexpected character `0` (U+0030) before name, expected a character that can start a name, such as a letter, `$`, or `_` More information: https://mdxjs.com/docs/troubleshooting-mdx
Is repowise free to use?
Yes. The core engine is AGPL-3.0, runs 100% locally, and the first index needs no API key. Paid hosted tiers start at $15/month for Pro (private repos, hosted re-indexing) and $20/seat/month for teams (3-seat minimum). Enterprise pricing covers on-premises, air-gapped, or commercially licensed deployments.
Which AI editors and agents does repowise work with?
repowise exposes your codebase over the Model Context Protocol, so it works with Claude Code, Cursor, Cline, Codex, and any MCP-compatible client. One index serves every agent; no separate integrations are needed per tool.
Does the code health score actually predict bugs?
Yes, and it is externally validated. Across 21 repositories and 9 languages, the cross-project ROC AUC is 0.74 (confidence interval 0.683 to 0.787, with up to 0.90 per repo). Ranking by repowise health surfaces 2.3x the defects of a leading commercial code health tool under the same review budget on the same 2,770 files. The benchmark harness and corpus are published.
Similar open-source tools#
Local Deep Research
Your AI research assistant, fully local and encrypted.
OpenFang
Open source Agent OS built in Rust with autonomous agents
r3
Minimalist Forth language for games and systems programming
mex
Persistent memory and code graph for AI coding agents
Switchyard
LLM proxy with API translation and multi-backend routing
ragflow
Open-source RAG engine with document parsing and agent orchestration