Scout setup recommendations

Scout works from any repository with no project to load. These settings help an agent reach for it.

Semantic search — off by default, and how to turn it on

Scout matches by meaning as well as text, so an agent can ask “where do we handle an inbound request?” and get the request handler rather than files containing those words. This is off by default. Enabling it embeds your whole workspace in the background, and that cost grows with the repository. To turn it on, add --semantic to the args in your MCP client config, or put the block below in .glider/scout/config.toml. On its next start Scout downloads the embedding model once (about 86 MiB), verifies it against a checksum built into the binary, and stores it in ~/.glidermcp/models/ so every repository on the machine shares one copy. Searching works throughout — semantic queries fall back to lexical until the tier is ready, and server_status reports which state it is in. Left off, Scout downloads nothing, so --no-telemetry alone makes it reach the network not at all.

# .glider/scout/config.toml
[semantic]
enabled = true

Install globally so it is on PATH

One global install serves every client and project without resolving the package at each launch. Install it globally, or set SCOUT_BINARY_PATH. An npx-only MCP entry runs Scout for that process and leaves no binary on PATH.

npm install -g @glidermcp/scout
curl -fsSL https://glidermcp.com/install.sh | sh   # Linux/macOS, no Node
brew install glidermcp/tap/scout                   # macOS/Linux

Agent system prompt instruction

Add this product-specific instruction to your MCP client system prompt or project instructions after connecting Scout.

Use the scout mcp find tool as the primary workspace search instead of shell grep, rg, or find. Strict modes (literal, regex, word) are exhaustive; fuzzy is ranked. If glider (C#) or tglider (TypeScript/JavaScript) is installed, prefer it for its language; use scout for every other language and as the universal fallback. Text search is the exception: tglider has no text-search tool, so literal and regex hunts in TypeScript and JavaScript belong to scout even when tglider is installed, and glider searches only the documents its workspace loaded.

Scope it to the project

Scout resolves its workspace root from the working directory, or pin it with --root <dir>. Prefer a project-scoped MCP config (Claude Code .mcp.json, Cursor .cursor/mcp.json, VS Code .vscode/mcp.json, Gemini CLI .gemini/settings.json, OpenCode opencode.json, Pi .mcp.json) so it indexes the repository you are working in. Scout works in any repository, so Claude Desktop’s single global config is also fine.