Claude Code

SourcePrep is purpose-built for Claude Code. Six MCP tools, auto-approve safety, skills integration, and client-aware content delivery.

Claude Code setup guide Request Beta Access

Why Claude Code + SourcePrep

Claude Code is the most capable AI coding CLI — it reasons, plans, and executes across an entire codebase. But even the best model is limited by what context it can see. SourcePrep gives Claude Code structural codebase intelligence via MCP: the full module graph, import chains, hub files, design decisions, and semantic search. Together, Claude Code doesn't just read files — it understands architecture.

Structural Awareness

Claude sees the full dependency graph, hub files, and import chains — not just flat file contents.

Agent-Native

SourcePrep is designed for autonomous agentic loops — low noise, high signal, source-cited.

Zero Ramp-Up

Call prep once at the start of any task. Structural context lands in seconds.

Six Tools at a Glance

SourcePrep registers six MCP tools with Claude Code. Each is read-only and purpose-built for a specific type of codebase intelligence.

prep

Ambient structural context, no query needed

prep_search

Semantic search with intent classification (7 intents auto-detected)

prep_impact

Blast radius analysis before changes

prep_audit

Enrich linter findings with structural context

prep_observe

Persistent cross-session memory

prep_concepts

Business rationale and design decisions

Auto-Approve & Skills

All six SourcePrep tools are read-only and completely safe to auto-approve. No writes, no side effects, no shell commands. SourcePrep generates the auto-approve config and a skills file for your project automatically — so Claude Code can call SourcePrep without a permission prompt on every invocation.

// .claude/settings.json

{

"permissions": {

"allow": ["mcp__prep"]

}

SourcePrep also generates a .claude/skills/ file that tells Claude Code when and how to call each tool — reducing friction and ensuring tools are used at the right moment in agentic workflows.

AGENTS.md Generation

SourcePrep auto-generates an AGENTS.md file at the root of your project. This file is the bridge between static project knowledge and live queryable intelligence — it tells Claude Code about the MCP tools available, provides a codebase atlas, lists hub files and entry points, and maps the full module structure.

Codebase AtlasAuto-generated structural snapshot: stack, workspace segments, entry points, cross-cutting concerns

Hub FilesFiles with the highest dependency fan-in — the places most likely to cause cascading changes

Module MapWorkspace boundaries and the symbols each segment exposes to others

Tool InstructionsFront-loaded, aggressive instructions telling Claude Code to call prep immediately at task start

The file is regenerated whenever SourcePrep rebuilds the index. User additions outside the managed block are preserved across regenerations.

Client-Aware Delivery

SourcePrep detects it's talking to Claude Code and tailors content delivery accordingly. Claude Code gets compact, high-density responses optimized for its context window — not the verbose output suited for a human-facing dashboard. Token budgets, response format, and tool instructions are all tuned per-client automatically.

Claude Code

Compact response format
Optimized token budgets
Claude-specific tool instructions
Skills + auto-approve config generated

Other Clients

Codex: AGENTS.md-shaped output, mcpServers config
Cursor: role-based atlas projection
Windsurf: verbose structural context
Generic MCP: standard response

Claude Code setup guide