One Server, Every Editor

IDE Integrations

SourcePrep speaks the Model Context Protocol. Any editor or agent that speaks MCP connects to the same daemon with the same config — no per-editor forks, no bespoke plugins.

One Server, Every Editor

Run prep serve once. Every MCP-aware agent in your workflow — CLI or IDE — reaches the same daemon and gets the same structural intelligence. No per-editor plugins, no vendor lock-in, no duplicated indexes.

Single Command

One daemon serves every connected agent simultaneously

MCP Standard

Works with any tool that implements the Model Context Protocol

No Lock-in

Switch tools freely — your index and context travel with you

Shared MCP config

"prep": {
  "command": "prep",
  "args": ["mcp"]
}

This is the common core. Every editor nests it slightly differently (servers vs mcpServers vs context_servers) and stores it at a different path. SourcePrep auto-generates these files when it detects a supported client.

MCP config setup guides

CLI Agents

Terminal-first coding agents see the biggest lift from structural context because they have no sidebar, no open tabs, no visible file tree to orient themselves. Claude Code is the primary target — the tool dispatch, auto-approve hints, and AGENTS.md generation are all shaped around it — but the MCP server works the same for Codex, Gemini CLI, Qwen Code, and other MCP-aware CLIs. (Codex is also a first-class AGENTS.md consumer, so the same generated file unlocks it.)

claude — my-project
Claude Code
Anthropic
Primary target — deepest hooks
OpenAI Codex
OpenAI
Native MCP + reads AGENTS.md
Gemini CLI
Google
Native MCP support
Qwen Code
Alibaba
Native MCP support
Any MCP CLI
Aider, Amp, Zed terminal, and others
Claude Code deep dive

Agentic IDEs

The big agentic editors all speak MCP natively — Cursor, Windsurf, Antigravity, and VS Code (via GitHub Copilot), plus Zed. The setup story is the same everywhere: drop the SourcePrep MCP block into the editor's config, restart, done. What differs is the editor itself, not the integration. MCP-aware extensions inside VS Code (Cline, Roo, CodeGPT) all work too — they read the same daemon.

Welcome×
1
// Ask the agent to open or edit a file.
SourcePrep Projectmain*TypeScript
AGENT
Ask the agent anything...
Cursor
Cursor
Primary IDE target — native MCP
Windsurf
Codeium
Native MCP
Antigravity
Google
Native MCP
VS Code + Copilot
Microsoft
MCP via GitHub Copilot
Zed
Zed Industries
Native context_servers
VS Code extensions
Cline · Roo · CodeGPT
MCP-aware extensions inside VS Code

Client-Aware Delivery

The setup is uniform, but the output isn't. SourcePrep reads the MCP clientInfo handshake and tunes the response per client — format density, context budget, and which hints to attach. Agents get shaped output without any per-editor configuration.

Claude CodeCompact format, auto-approve hints, skills context
OpenAI CodexAGENTS.md-shaped output; MCP block routed via mcpServers
CursorRicher format with inline code blocks and annotations
WindsurfBudget tuned to Windsurf's context window
Integration setup guides