Juno: local content indexing lifts Claude accuracy 91%→97% at 1/3 the tokens
Stock_Custard427 · reddit · 2026-11-19
The author built Juno because Windows search only matches filenames. It indexes file contents — PDFs (scans via OCR), PPT, Word, Excel, text in photos — with Ctrl+Alt+L global search by keyword or meaning, jumping to the exact page/slide. An MCP server connects it to Claude Code, Codex, Cursor, and VS Code so agents can search docs and code by behavior, symbol, and callers. In a test (109 questions, 1,080 files), Claude Sonnet 5.5 hit 97% accuracy with Juno vs 91% with its native Glob/Grep/Read/shell, using about a third of the tokens. Stack: Python, SQLite FTS5 + EmbeddingGemma 2 on bundled llama.cpp, merged with reciprocal-rank fusion. Windows-only, open source.
More from coding & agent
- New MAC Bench finds frontier agents coordinate poorly: 8-agent teams win by exploration, not teamwork — AlexGDimakis · 2026-11-19
- SimplePost turns ChatGPT into a social media publisher with in-chat previews — haltakov · 2026-11-19
- 10 Free Repos That Teach AI Agent Harness Engineering: Tools, Memory, Rules — blaizedsouza · 2026-11-19
- LangSmith ships coding agent ROI dashboard: $34.65 average cost per merged PR — LangChain · 2026-11-19
- LangSmith links coding agent spend to merged PRs to measure ROI — LangChain · 2026-11-19
- Pichai demos Gemini 3: napkin sketches become websites, images become board games — heyneighbor · 2026-11-19