Repowise indexes a repository on your machine into one local index of its call graph, git history, tests and coverage, generated docs, and design decisions, then answers questions from that index: what calls this, what breaks if it changes, what is dead, what to fix first, and why it was built this way. Coding agents query the index over the Model Context Protocol (MCP), and people reach the same index through a dashboard, a CLI, a VS Code extension, a pull-request bot, and CI gates. The index updates incrementally after each commit.
The graph, risk, health, test, dead-code, and PR review layers make no LLM calls and need no API key. Model-written wiki prose, decision extraction from agent transcripts, and refactoring code generation are optional, run against a provider you configure with your own key or offline through Ollama, and can be switched off. Raw source is parsed in memory and never persisted. It runs self-hosted from the open-source package, as a commercial self-hosted deployment inside a VPC or air-gapped network (Postgres with LanceDB or pgvector), or as a hosted service.
| Layer | What it contributes |
|---|---|
| Graph | File and symbol dependencies across 26 AST-parsed languages, confidence-stamped call resolution, communities, centrality, cycles, and execution flows |
| Git history | Hotspots, ownership, co-change, bus factor, and bug-fix history |
| Docs | A wiki for every module and file, hybrid search, and markdown checked against the tree for claims the code no longer supports |
| Decisions | Architectural rationale from ADRs, inline markers, commits, PRs, and agent sessions, each claim traced to evidence |
| Health and change | 53 deterministic detectors across defect risk, maintainability, and performance, change risk, test impact, dead code, and refactoring plans |
Claude Code and Codex CLI are at the Full integration tier: MCP tools, skills, slash commands, a managed instructions file, hook-level interception of tool calls, and transcript mining after the session. VS Code, Cursor, OpenCode, Hermes, GitHub Copilot CLI, and Kiro are at the Good tier (MCP tools plus an instructions file or skills), and any host that reads an mcpServers config, such as Cline, Windsurf, Zed, or Gemini CLI, takes the server from a printed config snippet. Ten task-shaped MCP tools ship by default with seven more opt-in, and every response carries the indexed commit, the index age, and a stale warning.
| Rung | Languages | What you get |
|---|---|---|
| Full (13) | Python, TypeScript, JavaScript, Svelte, Vue, Java, Kotlin, Go, Rust, C++, C#, Scala, Ruby | Symbols, import resolution, call graph, heritage, docstrings, framework edges, health markers |
| Good (11) | C, Swift, PHP, Dart, Object Pascal, COBOL, GDScript, VB.NET, Elixir, F#, Objective-C | The same minus the full health suite |
| Partial (2) | Luau, Razor | Symbols and some call edges |
| Lightweight (6) | Clojure, Haskell, Lean 4, Erlang, HTML, QML | File-to-file import graph |
| Structural (8) | R, Zig, Julia, Elm, OCaml, Crystal, Nim, D | Git history only |
Published measurements: on 42 sealed ContextBench instances the get_answer tool found 0.876 of the files a fix needed against 0.610 for the next tool, and over 43 django questions it cut an agent's output tokens by 31.6% while using 3.8 tool calls instead of 7.2. Call-graph precision against the Go and TypeScript compilers' own resolution runs 0.976 to 0.995, while about 14% of hand-graded call edges are wrong. At a 20% review budget on 2,770 files the health score surfaced 0.173 of defects against 0.074 for CodeScene. The largest repository indexed is dotnet/runtime, 58,924 files in 120 minutes at 11.7 GiB peak memory.
Features
- Call graph: file and symbol dependencies across 26 AST-parsed languages, each call edge stamped with how it was resolved and a confidence from 0.95 (same file) to 0.50 (repo-wide name match), plus communities, cycles, and traced execution flows
- Change risk: symbol-level blast radius for a diff, co-change partners missing from it, the tests that reach it, and a 0-10 rank against the repository's own recent commits, returned as a directive an agent can act on
- Test intelligence: the tests a diff exercises, read from a coverage report (LCOV, Cobertura, Clover, JaCoCo, Go) when one exists and inferred from the call graph when not, with patch and branch coverage gates in CI
- Code health: a 1 to 10 score per file from 53 deterministic detectors split into defect risk, maintainability, and performance, validated against the repository's own bug history, with a ranked Fix first queue
- Dead code and refactoring: unreachable files, unused exports, and unused packages by confidence tier, and refactoring plans (Extract Method, Extract Class, Move Method, Split File, Break Cycle) naming the symbols that move
- Docs and decisions: a structural wiki per module and file with optional model-written prose, markdown checked for drift against the tree, and decisions mined from ADRs, WHY comments, commits, PRs, and agent sessions
- Ownership and security: owners, bus factor, knowledge-loss risk, and suggested reviewers, plus secrets and risky calls found in the working tree and full git history with a pre-commit check and a CI gate
- MCP surface: ten task-shaped tools (get_overview, get_answer, get_context, get_symbol, search_codebase, get_risk, get_change_risk, get_why, get_dead_code, get_health) with seven more opt-in, each response carrying the indexed commit and its staleness
- Agent hooks and transcripts: hooks push the governing decision, bug-magnet warnings, and a session briefing into Claude Code and Codex, and optional transcript capture turns repeated corrections into tracked decisions
- Generated instructions files: writes and maintains the CLAUDE.md and AGENTS.md instructions files from the index, and a distill command shrinks command output reversibly before an agent reads it
- Workspaces: HTTP, gRPC, event, socket, and data contracts matched across repositories, cross-repo blast radius, a breaking-change guard, architecture rules checked in CI, and one federated MCP server and dashboard
- PR bot and CI: a deterministic GitHub App review with zero LLM calls, a GitHub Action and GitLab template for the gates, SARIF and GitLab Code Quality output, and webhook, hook, watch, or polling auto-sync
