Miru: Semantic Code Search MCP Cuts Agent Cost 20%, Turns 17%
ManRowing · reddit · 2026-09-08
Takara.ai shares a one-month retrospective on Miru, their open-source MCP server for semantic code search. Agents burn tokens orienting via grep-read-grep loops; Miru matches on meaning, returning ranked results (path, line range, snippet) for plain-English queries.
Key points:
- Four tools (search, expand, findrelated, locate), stdio transport; miru install wires into Claude Code, Cursor, Copilot, VS Code, Codex and Kiro.
- Device login flow now replaces pasting API keys, reducing secret exposure.
- Experimental benchmark mode hit a wall: agent behaviour is non-deterministic — the same question yields different tool-call sequences per run, so single-run comparisons are noise; the author asks for advice on reproducible agent measurement.
- Numbers: across 41 questions on a large codebase, 3 runs each — 20% cheaper, 17% fewer turns, 11% faster vs grep. When the LLM knows the exact string, grep still wins.
- Transparent model: CLI/server are MIT, but semantic search depends on the paid ds1-code embedding ($0.12/M tokens, free tier), self-hostable in your own AWS account; no vector DB involved.
More from coding & agent
- Cursor silently bypasses MCP fetch tools in favor of its built-in browser — InsideDebt6345 · 2026-09-08
- François Fleuret: software will soon be AI-generated locally per user, ending broadcast distribution — francoisfleuret · 2026-09-08
- 150 open headcounts: backend teams scramble as scale-driven complexity explodes — AlchainHust · 2026-09-08
- Replit's free mode called its best move yet amid rising dev costs — Kyrannio · 2026-09-08
- Same Prompt, Three Tools: ChatGPT vs Google Stitch vs Figma Make Compared — Tegadesigns · 2026-09-08
- Dev costs creeping up: newer models make AI-assisted development pricier — Kyrannio · 2026-09-08