🔍 unumsearch
The detective for your code. Your AI agents search the same files all day; unumsearch keeps the clues.
The case in 30 seconds
An AI coding agent asks the same kind of question hundreds of times a day: where is this
function called? which files mention that setting? A plain search like rg answers by reading
every file, every time. Over dozens of checkouts that is gigabytes per question.
unumsearch works like a detective with a good notebook. It keeps a compact index of clues: every 3-letter piece of text and the files it appears in. A question becomes a set of clues, the index names the few suspect files that hold all of them, and only those are opened and checked line by line against the real file. Same answer as ripgrep, usually in milliseconds.
Watch the race

Live run, real timings: 12 search patterns our agents actually issued, over ~180 agent worktrees
(258k files), one engine at a time with a warm page cache. Nothing is replayed or sped up. Run it on
your own code with python3 bench/race.py --root . (see it yourself).

How the detective works
- Never silently stale. A file watcher notices every edit; changed files are searched directly
until the index catches up (usually well under a second). Every answer says whether it is
fresh. - Never guesses. The index only narrows which files to read; the answer always comes from
the real file contents, with the same file rules as ripgrep (
.gitignore, hidden files, binaries and size cap). - Never peeks at secrets.
.ssh/,.aws/,.env*,*.key,*.pem,secrets.jsonand other credential files are never indexed, whatever the configuration says. - Travels light. No GPU, no model, no network. One static binary; the index is a memory-mapped file the OS can page out.
The evidence
20 regexes mined from a week of real coding-agent tool calls, on the author's machine (2026-10-09). Honest notes:
- Keep the daemon running. The speed comes from a process that keeps the index open (the
daemon, MCP server or
rpc). A new CLI process per query is slower thanrgon a small repository (12 ms vs 9.8 ms). - Same answers. 19 of 20 queries returned exactly rg's files over the whole tree; the 20th
differs only by
.env.examplefiles, which unumsearch never indexes on purpose. - Branch switches. When
git checkoutrewrites 44,686 Linux kernel files, the index has caught up 6.4 to 7.5 s after the checkout began; until then answers sayfresh: false, so an agent can fall back to its own scan. 0 of 64,454 fresh answers differed from ripgrep.
Method, every table and the weak spots: docs/BENCHMARKS.md. What is guaranteed and how it is tested: RELIABILITY.md.
Try it in one minute
curl -fsSL https://github.com/corpunum/unumsearch/releases/latest/download/install.sh | sh
irm https://github.com/corpunum/unumsearch/releases/latest/download/install.ps1 | iex
unumsearch --root ~/src/myrepo index # build the index once
unumsearch search 'fn\s+main' ~/src/myrepo --text # ask a question
unumsearch --root ~/src/myrepo serve # keep it fresh + serve the local API
Static binaries for Linux (x86_64, aarch64), macOS (arm64, x86_64) and Windows (x86_64), each
verified against SHA256SUMS with a build-provenance attestation. Config file, systemd unit and
building from source: Reference → Install.
Hand it to your agent
Checked, per-tool setup: docs/integrations.
Every answer is JSON with backend (index/scan), covered, fresh, truncated and
complete, so a client knows exactly when to trust it and when to scan itself. Inside: per-unit
trigram shards with delta/varint postings and a next-byte mask, a regex-to-trigram planner, and
incremental rebuilds that re-read only changed files. Details: Reference → How it works.
- Index or scan secret locations (
.ssh,.gnupg,.aws,.config,.kube,.env*, keys, credential files), from any front-end. - Answer the HTTP API for directories outside its configured roots, or for foreign
Hostheaders. - Listen beyond loopback without a token (
auth_token/UNUMSEARCH_TOKEN). - Call a model or open an outbound connection.
Full list: Reference → Security.
Learn more
Reference (every option, API, FAQ) · Benchmarks · Reliability · Integrations · Changelog · Agent Skill
Apache-2.0 (includes the MPL-2.0 nucleo-matcher dependency) · built by corpunum · sibling project:
UnumPingPong (AI agents rally on your tasks; a referee decides who scores)