ngosangns Knowledge Base

Repowise

#technology#code-intelligence#memory

Updated: 2026-09-24

Source

Definition

Repowise is more than a code knowledge graph. It merges graph + git history + docs + "decisions" (ADR mining) + code health into a single index that serves both agents and humans (local dashboard, PR bot). Its tagline is "Understand your codebase without paying your agent to rediscover it." The graph layer covers 26 AST-parsed languages with call resolution, communities, and execution flows. The git layer adds hotspots, ownership, and bus factor, which static analysis cannot see. A risk command scores a change from 0 to 10, and a free PR bot comments on GitHub. The core runs without an LLM, so the analysis stays deterministic. It ships under AGPL-3.0 or a commercial license.

Key Metrics

  • Stars: ~6.7k (repo created 03/2026)
  • Forks: ~713
  • License: AGPL-3.0 or a separate commercial license

Installation & Quick Start

pip install repowise
cd /path/to/your/repo
repowise init --no-prose -y
repowise serve

init builds the graph, git, decisions, health, dead-code, and structural-wiki layers in one go and wires up Claude Code automatically. No API key is needed for the deterministic core.

10 MCP Tools (task-shaped)

Unlike other tools that are typically designed around entities (one file, one symbol) and force agents into sequential call chains, Repowise is designed around tasks: pass multiple targets in a single call and get full context back.

Index Layers

Layer Content
Graph File + symbol dependencies across 26 AST-parsed languages, call resolution with confidence scores, communities, centrality, cycles, execution flows
Git Hotspots, ownership, co-change, bus factor, bug-fix history — behavioral signals that static analysis cannot see
Docs Per-module/file wiki, incremental rebuild with freshness/confidence scores + hybrid search
Decisions Architectural rationale mined from 5 index-time sources + manual notes from humans/agents, every claim with an evidence trace
Code health 49 deterministic detectors (defect risk, maintainability, performance) with concrete refactor plans

Notable Features

  • Change risk: repowise risk main..HEAD scores 0-10 based on the repo's own commit distribution; PR mode returns directives (may_break, missing_cochanges, missing_tests, tests_to_run).
  • Test intelligence: repowise impacted-tests — finds which tests actually cover a file via the call graph, with no coverage report needed; it can ingest LCOV/Cobertura/Clover when available.
  • repowise distill <cmd>: compresses command output (pytest, git log, ...) before the agent reads it, preserving exit codes/errors; repowise expand <ref> restores the truncated parts.
  • Decision mining from transcripts: enable repowise decision source set session --on to read agent transcripts, detect repeated corrections ("use the shared HTTP client, not raw requests"), and turn them into decision records — transcripts never leave the machine, and --no-llm skips the model-call step.
  • PR Bot (GitHub App): one comment per PR, edited in place on every push, silent if the PR is "green" — includes symbol-level blast radius, missing tests, change risk, and a Check Run that can gate merges.
  • Auto-generates CLAUDE.md/AGENTS.md from the index, so it is still useful for agents without MCP support.

Strengths

Strength Description
More comprehensive than a "pure graph" Combines git analytics + code health + decisions, not just code structure
Zero LLM for the core Graph/risk/health/tests/dead-code/PR review are all deterministic; prose is optional
Free PR Bot Automated review right on GitHub, no agent configuration needed
Transparent benchmarks Published with samples, method, limitations, and even the rows where it loses

Weaknesses

Weakness Description
AGPL-3.0 Network copyleft — needs consideration if integrated into a closed product; a commercial license is available
Young Repo created 03/2026; community/watchers are still small relative to its star count
Python Not a native binary like codegraph/codebase-memory-mcp; requires a Python environment

When to Use

  • You need to understand codebase "health" (bug hotspots, who owns which file, architectural decisions), not just pure code relationships.
  • You want free automated PR review with no extra configuration.
  • You want to cut tokens/tool calls by compressing command output before the agent reads it.

References:

See also: Source Code Indexing Comparison · CodeGraph · Codebase Memory MCP