Repowise

What is Repowise?

Repowise is an open-source, evidence-backed codebase intelligence platform and Model Context Protocol (MCP) server engineered to solve the context inefficiency of autonomous coding agents. Instead of forcing agents like Claude Code, Cursor, Codex, or Windsurf to repeatedly grep, read, and rediscover codebase relationships on every prompt, Repowise pre-indexes abstract syntax trees (ASTs), dependency graphs, git co-change history, test reachability, and architectural decisions into a continuously updated local index. In empirical benchmarks, Repowise reduces coding agent output by 31.6% and shrinks retrieval context payloads by up to 97.2% (from 13,984 down to 393 tokens) while running completely locally with zero LLM API dependencies for its core graph analysis.

Key Features & Technical Capabilities

• Universal Codebase Dependency Graph: Constructs file- and symbol-level dependency graphs across 26 AST-parsed programming languages. Delivers confidence-stamped call resolution, community clustering, centrality scores, and execution flow tracking.
• Ten Task-Shaped MCP Tools (dev.repowise/repowise): Specifically designed around developer workflows rather than low-level data entities. Agents can query blast radius, caller graphs, and architectural context across multiple targets in a single MCP round trip.
• Zero-LLM Deterministic Engine: Dependency graphing, change risk scoring, test impact analysis, dead code identification, and code health scans run entirely on local CPU cycles without sending source code to external LLM providers or incurring API costs.
• Intelligent Command Distillation (repowise distill): Pre-processes noisy shell and test runner outputs (e.g. pytest, git log) before the agent reads them, eliminating 60% to 90% of irrelevant log tokens while preserving reversible reference anchors ([repowise#<ref>]).
• 51-Detector Code Health Scoring: Evaluates codebases on a 1–10 scale across defect risk, maintainability, and cross-functional performance bottlenecks (such as cross-file N+1 queries). Calibrated against empirical defect corpora, surfacing 2.3× more defects than CodeScene under identical review budgets.
• Symbol-Level Blast Radius & Change Risk: Evaluates pull request diffs on a 0–10 risk scale based on commit shapes, modified contracts, and historical bug magnets. Automatically determines the exact test subset exercised by a diff.
• Automated PR Bot (GitHub App): Adds a single, self-updating PR comment highlighting symbol-level contract modifications, outside callers, missing co-change companion files, and a visual blast-radius treemap.
• Local Visual Dashboard (repowise serve): Hosts an interactive local web UI featuring explorable dependency architecture graphs, interactive code-health heatmaps, natural language codebase chat, and auto-generated wiki documentation.

Benchmark: Repowise vs. Unassisted Agent Retrieval

CLI Quickstart & Usage

Install Repowise via PyPI and initialize your repository index in under two minutes:

# Install from PyPI
pip install repowise

# Initialize local codebase index (no API key required)
cd /path/to/your/project
repowise init --no-prose -y

# Start local web dashboard and MCP server
repowise serve

Inspect change risk and run impact-focused test suites before merging:

# Assess change risk on current branch against main
repowise risk main..HEAD

# Run only the test files directly exercised by the diff
repowise impacted-tests main..HEAD

# Distill test output to eliminate token waste for AI agents
repowise distill pytest

# View accumulated token and cost savings
repowise saved

Claude Code & MCP Integration

To connect Repowise to Claude Code, Cursor, or any Model Context Protocol host, configure your .mcp.json or Claude configuration file:

{
  "mcpServers": {
    "repowise": {
      "command": "repowise",
      "args": ["mcp"],
      "env": {}
    }
  }
}

Technical Specifications

Frequently Asked Questions