← ClaudeAtlas

bld-runtime-activate-mcpslisted

rm: Spawn an MCP server (jcodemunch, context-mode, shadcn) for one burst of code search or registry lookup, then shut it down. Use for "hard optimize", "token-optimize this search", or a big multi-file sprint.
mnguyenht/bld-package · ★ 1 · AI & Automation · score 64
Install: claude install-skill mnguyenht/bld-package
# bld-runtime-activate-mcps — burst-use the optimizer MCPs, then dip These are **unofficial third-party MCP servers** installed from GitHub (`jcodemunch-mcp` via pip, `context-mode` via npx). By default they stay **out of `.mcp.json`**, so they never auto-load and never ambiently route the codebase. This skill spins one up over stdio, fires the exact queries needed, prints compact results, and kills the process — seconds, not a session. Use it for a **big sprint** where you'd otherwise read many large files: pull the compact search results once, then work from them. ## Ground rules - **Do not hand-edit `.mcp.json` to make these permanent.** Always-on is a real option, but it belongs to `/bld-settings-mcp`, which refuses to delete entries it did not write. Editing the file directly loses that protection, and this skill keeps working either way. - Prefer **jcodemunch** — it's read-only code search. Reach for **context-mode** only if the task genuinely needs its shell/cross-tool features (its `ctx_execute` runs real commands with your logged-in `gh`/`aws`/`kubectl` — heavier trust). - **jcodemunch is blind to CSS** (`missing_extractors: css`) — its symbol/semantic tools return nothing for stylesheets. For CSS use `search_text`, or just read the file. Its strength is JS/TS/Python symbols. ## How to run it > **`python` is the Windows spelling.** macOS and most Linux distros ship the > interpreter as `python3` and have no bare `python` at all, so every command >