repobrain

Codebase intelligence that thinks ahead.

---

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.

It fixes every critical architectural flaw in repowise, then goes further.

---

Four Intelligence Layers

🔗 Graph Intelligence	⏱ Temporal Intelligence
tree-sitter + NetworkX dependency graph with PageRank centrality. Dynamic import hints for Django, pytest, and Node recover 20–40% missing edges that repowise silently drops.	Exponential decay scoring weights recent commits exponentially higher. A commit from yesterday matters far more than one from 2 years ago. Configurable half-life.

📚 Documentation Intelligence	💥 PR Intelligence
RAGAwareDocGenerator fetches dependency docs from LanceDB before every LLM call — not after. This is the #1 architectural difference from repowise.	PRBlastRadiusAnalyzer traces direct + transitive impact of every PR, scores risk 0–10, flags co-change pairs, and recommends reviewers.

---

Why repobrain Beats Repowise

#	Repowise Flaw	repobrain Fix
1	RAG context never used during generation — vector store populated but never queried	RAGAwareDocGenerator fetches dependency docs from LanceDB before every LLM call
2	25+ min initial indexing — no parallelism	7-stage async pipeline; parse in ProcessPoolExecutor, git + graph run concurrently
3	3 stores, no atomic transactions — 5–15% silent consistency failures	AtomicStorageCoordinator buffers + rolls back SQL, LanceDB, and NetworkX atomically
4	Hardcoded 500-commit limit	GitConfig.max_commits = 10_000 — fully configurable
5	Dynamic imports invisible (Django, pytest, importlib) — 20–40% missing graph edges	DjangoDynamicHints, PytestDynamicHints, NodeDynamicHints in HintRegistry
6	Incremental updates miss global percentile recalculation	upsert() always triggers PERCENT_RANK() window function refresh
7	No PR blast radius analysis	PRBlastRadiusAnalyzer + repobrain review <PR> + get_pr_impact MCP tool
8	Temporal blindness — 3-year-old commits weighted same as yesterday's	Exponential decay: score += exp(-ln(2) × age_days / halflife) × complexity
9	Zero cost visibility	TokenspyCostAdapter wraps every Anthropic call; repobrain costs CLI
10	Conservative dead code misses real candidates	Dynamic hint edges recovered; configurable sensitivity threshold

Full comparison →

---

What's New (Not in Repowise)

Tool	Description
get_pr_impact	Full blast radius before merge: direct + transitive files, risk score 0–10, co-change warnings, reviewer recommendations
get_knowledge_map	Knowledge silos, bus factor per module, onboarding targets for new engineers
get_test_gaps	Untested code ranked by combined risk score (hotspot × centrality)
get_security_hotspots	Auth handlers, input validation points, SQL surfaces — ranked by exposure

---

Installation

pip install repobrain

Requirements: Python 3.12+, Anthropic API key.

---

Quick Start

# 1. Configure
export ANTHROPIC_API_KEY=sk-ant-...

# 2. Index your repository (3–5 min for 1,000 files)
repobrain index /path/to/your/repo

# 3. Analyze a PR before merging
repobrain review 42

# 4. Start MCP server for Claude Code
repobrain serve

Add to your Claude Code MCP config:

{
  "mcpServers": {
    "repobrain": {
      "command": "repobrain",
      "args": ["serve", "--mcp-only"]
    }
  }
}

---

CLI Commands

Command	Description
repobrain index [PATH]	Index a repository (full or incremental)
repobrain review <PR>	PR blast radius: risk score, affected files, reviewer recommendations
repobrain serve	Start MCP server (+ optional GitHub webhook)
repobrain status	Temporal hotspot rankings and index health
repobrain query "<text>"	Natural language codebase search
repobrain costs [--since DATE]	Per-operation LLM spend breakdown

Rich progress during indexing:

[=====>     ] 47% │ Stage: Generating Docs │ Files: 234/500 │ Cost: $0.23 │ ETA: 4m 12s

---

MCP Tools (12 total)

Tool	Status	Description
explain_file	—	File docs with RAG-injected dependency context
explain_symbol	—	Symbol-level explanation
get_hotspots	—	Temporal decay–weighted churn hotspots
get_ownership	—	Temporal-weighted file ownership
get_dependencies	—	Import graph with dynamic hint edges
get_architectural_decisions	—	ADR search and retrieval
search_codebase	—	Semantic vector search
get_cochange_patterns	—	Temporal co-change analysis
get_pr_impact	NEW	Full blast radius for a PR
get_knowledge_map	NEW	Knowledge silos, bus factor, onboarding targets
get_test_gaps	NEW	Untested code ranked by risk score
get_security_hotspots	NEW	Auth/input/SQL risk surfaces

---

Language Support

Language	Parsing	Graph	Dynamic Hints
Python	✅ tree-sitter	✅ imports + calls	✅ Django, pytest, importlib
TypeScript	✅ tree-sitter	✅ imports + re-exports	—
JavaScript	✅ tree-sitter	✅ require + import	✅ Node.js patterns
Go	✅ tree-sitter	✅ package imports	—

Config files (Dockerfile, GitHub Actions, Makefile, requirements.txt) are indexed as nodes in the dependency graph even without a full parser.

---

Architecture

repobrain index /repo
      │
      ▼
┌─────────────────────────────────────────────────────┐
│            AsyncIndexingPipeline  (7 stages)         │
│                                                      │
│  1. Discovery    ──▶ file manifest to SQL            │
│  2. Parse        ──▶ ProcessPoolExecutor (CPU-bound) │
│  3. Graph Build  ─┐  concurrent                      │
│  4. Git Analysis ─┘  (asyncio.gather)                │
│  5. Embedding    ──▶ ThreadPoolExecutor + semaphore  │
│  6. RAG Doc Gen  ──▶ LanceDB deps fetched FIRST      │
│  7. Atomic Commit──▶ AtomicStorageCoordinator.txn()  │
└─────────────────────────────────────────────────────┘
      │
      ▼
┌──────────┐   ┌──────────────┐   ┌────────────────┐
│  SQLite  │   │   LanceDB    │   │ NetworkX Graph │
│ (files,  │   │ (embeddings, │   │ (dependency    │
│  metrics)│   │  docs)       │   │  graph)        │
└──────────┘   └──────────────┘   └────────────────┘

Atomic transactions across all three stores:

async with coordinator.transaction() as txn:
    txn.pending_sql.append(...)
    txn.pending_vectors.append(...)
    txn.pending_edges.append(...)
# On exception: SQL rollback + vector delete + graph node removal — guaranteed.

---

Privacy & Security

• Self-hosted — your code never leaves your machine
• BYOK — bring your own Anthropic API key
• No telemetry — zero usage data collected or transmitted
• HMAC-SHA256 webhook validation for GitHub integrations
• MIT license — no AGPL restrictions, use commercially without a commercial license

---

Screenshots

repobrain index	repobrain review

repobrain status	repobrain costs

Full interactive demo at pinexai.github.io/repobrain →

---

Configuration

# .env
ANTHROPIC_API_KEY=sk-ant-...

REPOMIND_DATA_DIR=~/.repobrain          # index storage location
REPOMIND_MAX_COMMITS=10000              # git history depth (repowise hardcodes 500)
REPOMIND_DECAY_HALFLIFE_DAYS=180        # temporal scoring half-life
REPOMIND_GENERATION_CONCURRENCY=5       # parallel doc generation workers
REPOMIND_MCP_PORT=8766                  # MCP server port
REPOMIND_WEBHOOK_PORT=8765              # GitHub webhook port
REPOMIND_WEBHOOK_SECRET=your-secret    # HMAC-SHA256 webhook validation

---

Documentation

Full docs at pinexai.github.io/repobrain

Section	Link
Installation & Quick Start	Getting Started
CLI Reference	CLI Docs
MCP Tools Reference	MCP Docs
Architecture Deep Dive	Architecture
Migrate from repowise	Migration Guide
repobrain vs repowise	Comparison
Python API Reference	API Docs

---

License

MIT — see LICENSE.

Built by pinexai. Not affiliated with repowise.