Coding agents keep quoting PDFs and Word files they can't actually open. So I tested it. I asked an Opus agent 127 questions about 53 public documents with only Grep and Read, and 41 of its quotes weren't anywhere in the file it named. With Semlith it found the right passage 98% of the time and made up none.

Semlith is a local index of your code and documents that AI agents search over MCP instead of grepping and reading whole files. Claude Code, Codex, Cursor and other MCP clients get back the exact file and line. It reads PDF, Word, PowerPoint and Excel, keeps a code graph for callers and what breaks if something changes, and re-indexes as files change.

On SWE-bench Verified, the file the fix edits is in Semlith's top 10 for 76.8% of 500 issues, against 47.0% for grep then read. That's retrieval only. It doesn't help everywhere either. On single-repo code with a strong model it made no difference, 92% with or without it. Every benchmark has the command to rerun it.

It's a single Rust binary, free and open source under Apache-2.0, with every feature. No API keys, and nothing leaves your machine. It runs on macOS (Apple silicon), Linux and Windows, and one command installs it and sets it up for the agents it finds. Setup is tested for 12 clients.

If your agents also run in the cloud or in CI, or you want one index for a whole team, Semlith Cloud is the paid, hosted version. Pro starts at $5 a month on the founding price, with a 14-day trial and no card. You don't need it if all your agents run on your laptop.

I've been building it on my own since July.