Give your AI access to the epistemic context it needs to understand your codebase.
An MCP server that maps your entire codebase — imports, call chains, symbol hierarchies — and delivers deep structural intelligence to Claude Code, Antigravity, Cursor, Paperclip and any MCP-compatible tool.
DocumentationWhat is SourcePrep?
SourcePrep is a structural codebase intelligence MCP server for AI coding agents. It builds a code graph using Rust and tree-sitter to map imports, call chains, and symbol hierarchies — then delivers that deep structural context directly to Claude Code, OpenAI Codex, Cursor, Windsurf, and any other MCP-compatible tool. Connect your preferred LLM — Kimi 2.6, local models, or frontier APIs.
Prep the context before the AI call
Run prep before grep
Your AI agent orients itself by opening files and running grep — one guess at a time. SourcePrep is the prep step that hands it a structural map of the codebase up front, so every answer starts from a full picture instead of a keyword match. The payoff: 3–20× more signal per token than dumping whole files into the prompt.
The Problem: Naive search finds "Where" and has to figure out the rest
Large repos with 100k files with thousands or many large markdown docs can be challenge. Basic vector search grabs random files without understanding how code connects, flooding your agent's context window with noise and causing hallucinations.
The Fix: Structural Tracing + Enrichment
SourcePrep's Rust engine traces your entire codebase — mapping imports, call graphs, symbols, and hierarchies. Then it iterates though enriching it's understanding. It delivers structurally-aware context precisely scoped an agent's request or for each agent's role.
The Result: Agents That Understand Architecture and Concepts
Connect the SourcePrep MCP server to Claude Code, Antigravity, or Cursor and watch your AI instantly understand the codebase. By providing the "what, why, when, and how", agents get the structural awareness they need to make grep more stategic, your audit infomed or design more structually aware.
See The Tools
Six MCP tools. Deep codebase intelligence.
Connect SourcePrep once to Claude Code, Codex, Cursor, Windsurf, or any other MCP-compatible tool and your AI gets structural awareness, semantic search, blast radius analysis, audit enrichment, persistent memory, and recorded design rationale.
Integrations
Works with the editors and agents you already use
One MCP server, every client. Connect SourcePrep once and any MCP-aware tool picks it up.
Capabilities
Built for large codebases and sprawling doc trees
Structural Code Graph
Goes beyond vector search. A Rust engine traces your codebase to map imports, call graphs, and symbol hierarchies — giving agents the what, why, when, and how.
Graph Enrichment Pipeline
A multi-stage pipeline deepens understanding over time — from Rust parsing to deep LLM reasoning to architectural synthesis. Supports swarm mode for parallel LLM processing.
Audit Enrichment
Supercharge your existing linters. Feed findings from ruff, semgrep, or CodeQL into SourcePrep and get them back enriched with structural context — dependent count, hub status, risk score, and related concepts. Outputs SARIF for CI integration.
Role-Aware Context
Different agents need different slices. A security reviewer sees auth boundaries; a UI agent sees design tokens. SourcePrep shapes context delivery around the role of the agent asking.
Atlas Routing
Your codebase is mapped into subsystem segments at build time. When an agent asks for context, the Atlas routes the query to the correct architectural neighborhood before search begins.
Smart Context Compression
Intelligently shrinks the payload without losing structural integrity. Full source for top results, signatures for mid-relevance — delivering 3–20× more signal per token.
Native MCP Server
SourcePrep is an MCP server. Connect it to Claude Code, Antigravity, Cursor, VS Code, or any tool that supports the Model Context Protocol.
Persistent Agent Memory
Observations about decisions, bugs, and patterns persist across sessions — linked to specific files. When those files change, linked observations are flagged [STALE] so your AI knows to re-evaluate.
Concept System
Record the "why" behind your architecture — business rationale, design decisions, constraints. Concepts have testable assertions that auto-generate immune system rules, catching violations before they ship.
Built for professionals who take their code seriously
Got questions? Read the FAQ →