About codebase-synapse
codebase-synapse indexes a repository into a local knowledge graph, giving coding agents search, dependency analysis, and architecture information. It stores the index in SQLite and includes Git history analysis and cross-session memory tools.
Key features
- 01Semantic, full-text, and hybrid code search
- 02Traverse call graphs and dependencies
- 03Analyze impact and architectural boundaries
- 04Inspect Git history and change coupling
- 05Use persistent and session memory tools
- 06Block high-impact file writes with codebase-guard
Use cases
- 01Find callers and callees of a function
- 02Assess impact before editing a file
- 03Trace file history and co-changes
- 04Prepare code-edit context for an agent
MCP support
The stdio MCP server exposes code search, graph, architecture, Git history, and memory tools to MCP clients. The index is built separately with a CLI command; the server reads the existing local index.
Configuration
Local server config is the same for every client. Copy it into your client’s MCP config file.
Tools
Local servers expose tools only once running on your machine, and the maintainer has not provided a tool list.
Details
From the project README🧠 Codebase Synapse
Give your AI agent a deep understanding of your entire codebase.
An MCP server that indexes your codebase into a local knowledge graph with 52 AI tools — semantic search, call-graph traversal, git archaeology, blast-radius analysis, and more.
Works with Claude Code · Cursor · Windsurf · Zed · Any MCP client
The Problem
AI coding agents (Claude, Cursor, etc.) are powerful — but they work with limited context. They can only see the files you open or feed them. Ask about call chains, architectural patterns, or blast-radius impact, and they guess.
Codebase Synapse fixes this. It indexes your entire repo into a local knowledge graph stored in SQLite, giving your AI agent real answers backed by structural analysis — not hallucinations.
How it works
graph LR
A["📁 Your Codebase"] -->|Tree-sitter| B["🧩 Parser"]
B -->|Symbols & Edges| C["🕸️ Knowledge Graph"]
C -->|SQLite| D["💾 Local DB"]
D -->|52 MCP Tools| E["🤖 AI Agent"]
style A fill:#1a1b26,stroke:#7aa2f7,color:#c0caf5
style B fill:#1a1b26,stroke:#bb9af7,color:#c0caf5
style C fill:#1a1b26,stroke:#9ece6a,color:#c0caf5
style D fill:#1a1b26,stroke:#e0af68,color:#c0caf5
style E fill:#1a1b26,stroke:#f7768e,color:#c0caf5
⚡ Quick Start
1. Index your project (CLI, Zoekt-style)
codebase-synapse index /path/to/your/repo
Indexing runs as a separate CLI command and writes into ~/.codebase-synapse/codebase.db. The first index is the slow one; re-running it only re-indexes changed files. The MCP server never indexes — it only reads the pre-built index.
2. Start the MCP server
npx codebase-synapse
That's it. No Docker. No database setup. No cloud. The server starts via stdio and serves the pre-built index to your MCP client.
Configure your client
Claude Code / Claude Desktop
Add to your claude_desktop_config.json or mcp_servers.json:
{
"mcpServers": {
"codebase-synapse": {
"command": "npx",
"args": ["-y", "codebase-synapse"]
}
}
}
Cursor
Add to your .cursor/mcp.json:
{
"mcpServers": {
"codebase-synapse": {
"command": "npx",
"args": ["-y", "codebase-synapse"]
}
}
}
Other MCP clients
Any MCP client that supports stdio transport works. Just point it at:
npx -y codebase-synapse
✨ What It Does
🔍 Search & Discovery
| Tool | Description |
|---|---|
semantic_search | Vector similarity search using local embeddings (all-MiniLM-L6-v2) |
search_code | Full-text search across code (FTS5 + BM25 ranking) |
search_symbol | Find functions, classes, types by name or pattern |
hybrid_search | Combined semantic + lexical search with RRF fusion |
find_similar | Find structurally similar code using MinHash + LSH |
find_symbol_everywhere | Locate a symbol across all indexed projects |
🕸️ Knowledge Graph
| Tool | Description |
|---|---|
get_callers / get_callees | Navigate the call graph in either direction |
get_imports / get_dependents | Trace dependency chains |
impact_analysis | Compute blast radius before editing a file |
find_path | Find the shortest connection between two symbols |
find_dead_code | Detect unreachable functions and unused exports |
get_pagerank | Identify the most critical nodes in your architecture |
query_graph | Run Cypher-like queries against the knowledge graph |
🏗️ Architecture
| Tool | Description |
|---|---|
get_architecture | Full project architecture overview (languages, entry points, hotspots) |
get_file_structure | Directory tree with symbol annotations |
project_overview | High-level summary with key metrics |
get_route_map | Extract HTTP routes and their handler mappings |
suggest_boundaries | Detect module boundaries via Leiden clustering |
check_boundaries | Validate cross-module dependencies against defined boundaries |
get_clusters | Community detection across the codebase |
🔬 Git Archaeology
| Tool | Description |
|---|---|
git_archaeology | Deep history analysis of a file (authors, churn, evolution) |
get_hotspots | Files with highest complexity × change frequency |
detect_change_coupling | Files that always change together |
get_recent_semantic_changes | Semantically meaningful recent changes |
index_git_history | Build temporal analysis from git log |
🧠 Memory & Context
| Tool | Description |
|---|---|
memory_store / memory_search / memory_list | Persistent notes, facts, and decisions across sessions |
session_remember / session_recall | Short-term memory within a session |
get_context | Budgeted context preparation for AI agents |
get_edit_context | Focused context for a specific file edit |
get_working_set | Recently accessed and modified files |
manage_adr | Architecture Decision Records management |
🛡️ Codebase Guard
Included as a bonus: codebase-guard is a PreToolUse hook for Claude Code that blocks writes to high-impact files until the agent runs impact_analysis first.
It uses PageRank scores and blast-radius data from the knowledge graph to identify architectural hubs. No more accidental edits to core files.
🔧 Technical Details
| Language | Rust (compiled native binary) |
| Transport | MCP stdio (JSON-RPC) |
| Storage | SQLite (WAL mode, zero config) |
| Embeddings | all-MiniLM-L6-v2 via Candle (offline, local, lazy-loaded) |
| Parsing | Tree-sitter (10 languages) |
| Distribution | npm with prebuilt binaries (Windows, macOS, Linux × x64, arm64) |
Supported Languages
Rust · Python · TypeScript · JavaScript · Go · Java · C# · PHP · C · C++
Architecture
src/
├── parser/ # Tree-sitter parsing & entity extraction
├── graph/ # Knowledge graph, PageRank, Leiden clustering
├── indexer/ # Repository indexing pipeline
├── search/ # BM25 full-text + vector cosine + hybrid RRF
├── embedding/ # Candle-based local embeddings (feature-gated)
├── memory/ # Persistent & session memory stores
├── mcp/ # MCP protocol transport + 52 tool handlers
├── git/ # Git archaeology, intent classification, hotspots
├── context/ # Budgeted context preparation for AI
├── cypher/ # Nom-based Cypher parser → SQL CTE planner
├── similarity/ # MinHash + LSH structural similarity
├── semantic/ # Multi-signal scoring (tokens, directory, AST)
└── cli/ # Interactive TUI installer + artifact export/import
🤝 Contributing
Contributions are welcome! The project uses standard Rust tooling:
# Run tests
cargo test
# Lint
cargo clippy -- -D warnings
# Format
cargo fmt --all
📄 License
Comments
New to MCP?
From what the protocol is to installing your first server, these answers cover most first questions.
An open protocol that defines how AI applications connect to external tools and data in a uniform way. Clients talk to the model; MCP servers provide tools, resources and prompt templates. They exchange standard messages, so one server can be reused by many clients.