{
  "$type": "site.standard.document",
  "bskyPostRef": {
    "cid": "bafyreidgbzoicf5tjfagxus6qmsf7mb2de7hbzapo4o4tshj6myfl4kayu",
    "uri": "at://did:plc:2u26gaflouttm3uj6jkgroyz/app.bsky.feed.post/3mhtpgijyqsv2"
  },
  "coverImage": {
    "$type": "blob",
    "ref": {
      "$link": "bafkreid2sjlvru77zgg5d6nd7hbqy6lscfpw2ttpud7gjvivnqg4ytzi2e"
    },
    "mimeType": "image/png",
    "size": 2281209
  },
  "description": "By default, Repo Prompt installs its MCP server for Claude Code per-project via .mcp.json files. Here's how I made it available globally in every Claude Code session.",
  "path": "/til/2026/03/24/how-to-install-repoprompt-globally-in-claude-code/",
  "publishedAt": "2026-03-24T23:16:51.000Z",
  "site": "https://danielraffel.me",
  "tags": [
    "RepoPrompt",
    "AST-aware",
    "Agent Mode",
    "Claude Code"
  ],
  "textContent": "RepoPrompt is a context-focused IDE with agent tooling designed for reliability and token efficiency. I use it with agents like Claude/Codex when working with monorepos, large or familiar codebases, and projects that have become complex. In practice, it consistently improves my codegen success rate. It's been described to me as a:\n\n> • highly context-efficient agent tooling with AST-aware codemaps that let you target specific functionality precisely,\n> • a context builder that assembles deep, structured prompts from those tools,\n> • a chat/oracle layer that uses that context for extended analysis.\n>\n> The core issue is that agents rarely have enough context to see the full picture, so they mix research, planning, and review instead of treating them as separate steps.\n>\n> RepoPrompt functions as a context pump, applying context engineering to generate highly efficient analysis prompts.\n\nBy default, its MCP server for Claude Code is configured per-project via `.mcp.json` files. But if you want it available in every Claude Code session, regardless of which directory you're in, you can install it globally.\n\n## Agent Mode Is Amazing\n\nI run Repo Prompt in Agent Mode with Claude connected via MCP, so the agent can call Repo Prompt’s context engine natively inside the session. The result is that I rarely think about prompting. I just describe what I want, and the agent automatically pulls the right context or invokes Repo Prompt when needed. It feels like a clean separation of intent and execution, where I stay high-level and the system has a new super power that assists with discovery, context, and reasoning.\n\n## Prerequisites\n\n  * RepoPrompt installed (the macOS app)\n  * Claude Code installed\n\n\n\n## Background Context\n\nClaude Code has two config files that look like they'd be the right place for global MCP servers:\n\n  * `~/.claude/settings.json` — global settings (permissions, plugins, env vars)\n  * `~/.claude.json` — user-level config (MCP servers, preferences)\n\n\n\nDespite the naming, **MCP servers go in`~/.claude.json`**, not `~/.claude/settings.json`.\n\n## Step 1: Locate the RepoPrompt MCP Binary\n\nRepoPrompt ships an MCP server binary inside its app bundle. It also creates a convenient symlink during setup:\n\n\n    # The symlink (created by RepoPrompt's setup)\n    ~/RepoPrompt/repoprompt_cli\n    # Which points to the actual binary inside the app bundle\n    /Applications/Repo\\ Prompt.app/Contents/MacOS/repoprompt-mcp\n\nYou can verify this on your machine:\n\n\n    # Check the symlink exists and where it points\n    ls -la ~/RepoPrompt/repoprompt_cli\n\nIf the symlink doesn't exist, you can use the full app bundle path directly.\n\n## Step 2: Add RepoPrompt to ~/.claude.json\n\nOpen `~/.claude.json` and find the `\"mcpServers\"` object. Add the RepoPrompt entry alongside your other servers:\n\n\n    {\n      // ... other config (numStartups, autoUpdates, etc.) ...\n      \"mcpServers\": {\n        // ... your other MCP servers ...\n        // Add RepoPrompt globally — uses the symlink created by RepoPrompt's setup\n        \"RepoPrompt\": {\n          \"command\": \"/Users/YOUR_USERNAME/RepoPrompt/repoprompt_cli\",\n          \"args\": []\n        }\n      }\n      // ... rest of config ...\n    }\n\nReplace `YOUR_USERNAME` with your macOS username. Alternatively, use the full app bundle path if you prefer not to rely on the symlink:\n\n\n    {\n      \"mcpServers\": {\n        \"RepoPrompt\": {\n          // Direct path to the binary inside the app bundle\n          \"command\": \"/Applications/Repo Prompt.app/Contents/MacOS/repoprompt-mcp\",\n          \"args\": []\n        }\n      }\n    }\n\n## Step 3: Restart Claude Code\n\nRestart Claude Code, then run `/mcp` to verify. You should see:\n\n\n    RepoPrompt · ✔ connected\n\nThat's it. RepoPrompt's MCP tools: `context_builder`, `file_search`, `apply_edits`, `get_file_tree`, and more are now available globally in every Claude Code session.\n\n* * *\n\n## Agent Prompt\n\nAlternatively, skip the steps above and just use this prompt with Claude and let it do it for you:\n\n\n    Install RepoPrompt's MCP server globally for Claude Code. Here's what you need to know:\n    1. Find the RepoPrompt MCP binary. It's either at ~/RepoPrompt/repoprompt_cli (a symlink created by the app) or directly at /Applications/Repo Prompt.app/Contents/MacOS/repoprompt-mcp. Verify which path exists on my system.\n    2. Add it to ~/.claude.json (NOT ~/.claude/settings.json — MCP servers live in ~/.claude.json despite the naming). Find the \"mcpServers\" object and add a \"RepoPrompt\" entry with the binary path and empty args array.\n    3. Tell me to restart Claude Code and run /mcp to verify it shows as connected.",
  "title": "How to Install RepoPrompt Globally in Claude Code",
  "updatedAt": "2026-03-25T00:57:31.119Z"
}