Skip to content

repobrain

repobrain

Codebase intelligence that thinks ahead.

PyPI Python 3.12+ MIT MCP Compatible

Get Started    View on GitHub    vs. repowise


repobrain is a self-hosted MCP server for Claude that gives it deep, always-fresh understanding of your codebase — 10× faster indexing, RAG-aware documentation, PR blast radius, temporal hotspot scoring, and 12 MCP tools.


Live Demo


Six Core Capabilities

🔗 Graph Intelligence

tree-sitter parses Python, TypeScript, JavaScript, and Go into a NetworkX dependency graph. PageRank centrality identifies architectural hubs. Dynamic hint extractors (Django, pytest, Node) recover 20–40% of edges that static analysis misses.

⏱ Temporal Scoring

Exponential decay weights recent commits exponentially higher. A fix merged yesterday counts far more than code from 2 years ago. Half-life is configurable. Percentile ranks refresh on every incremental update.

📚 RAG Documentation

RAGAwareDocGenerator fetches dependency docs from LanceDB before every LLM call. This is the #1 architectural flaw in repowise — it populates the vector store but never queries it during generation.

💥 PR Blast Radius

PRBlastRadiusAnalyzer traces direct + transitive impact of every PR. Risk score 0–10. Co-change warnings for historically coupled files. Reviewer recommendations based on ownership.

🔒 Atomic Transactions

AtomicStorageCoordinator wraps every write across SQLite, LanceDB, and NetworkX in a single logical transaction. Any failure rolls back all three stores. No silent consistency failures (5–15% failure rate in repowise).

💰 Cost Tracking

TokenspyCostAdapter wraps every Anthropic call and records per-operation token spend. repobrain costs shows a breakdown by command, date, and model. Know exactly what your indexing costs.


Screenshots

repobrain index — 7-stage async pipeline

repobrain index

repobrain status — Temporal hotspot rankings

repobrain status

repobrain review 42 — PR blast radius analysis

repobrain review 42

repobrain costs — LLM spend by operation

repobrain costs

MCP Tools — 12 tools available in Claude Code

repobrain MCP Tools


Why repobrain?

After reverse-engineering repowise's architecture, we found 10 critical flaws. repobrain fixes every one:

# Repowise Flaw repobrain Fix
1 RAG context never used during generation Dependency docs fetched from LanceDB before every LLM call
2 25+ min initial indexing 7-stage async pipeline with ProcessPoolExecutor + concurrent git/parse
3 No atomic transactions across 3 stores AtomicStorageCoordinator rolls back SQL + LanceDB + NetworkX atomically
4 Hardcoded 500-commit limit Configurable GitConfig.max_commits = 10_000
5 Dynamic imports invisible (20–40% missing edges) Django, pytest, Node hint extractors
6 Incremental updates miss percentile recalculation upsert() always triggers PERCENT_RANK() window refresh
7 No PR blast radius analysis PRBlastRadiusAnalyzer + get_pr_impact MCP tool
8 Temporal blindness (old commits = recent commits) Exponential decay scoring
9 Zero cost visibility TokenspyCostAdapter + repobrain costs CLI
10 Conservative dead code detection Dynamic hint edge recovery

See the full comparison →


Get Started

pip install repobrain
repobrain index /path/to/your/repo
repobrain serve  # start MCP server for Claude Code

Returning repowise users: migrate in 5 minutes →

New users: Quick Start guide →