XERJ publishes v1.0.0-rc.96 of its local search engine

On 16 October 2026 at 14:51 UTC, XERJ published v1.0.0-rc.96. The changelog dates that section 15 October. The version string is a release candidate. GitHub’s release record does not flag it as a prerelease.

xerj-org/xerj was created on 30 June 2026. At 17:50 UTC on 16 October the GitHub API reported 6,704 stars and 339 forks. The license file at the tag is Apache License 2.0. This is the first account here, timed to rc.96. The repository has been public since June.

A Rust engine, with an Elasticsearch-shaped API

xerj.org describes one Rust binary under Apache-2.0, with an Elasticsearch-compatible HTTP API, and says Linux builds are static musl. llms.txt says xerj autoindex <folder> indexes mixed files, including PDF, DOCX, logs, and mail, and that xerj search prints a file, a line, and a passage. xerj init registers an MCP server in .mcp.json, writes a 10-line skill, and backs up files it edits.

GitHub’s language field for the repo is Rust. The tree at the tag also contains HTML, JavaScript, and Python (the site, demos, and benchmark scripts). A crates.io lookup for a crate named xerj found none. The pages describe API compatibility. They do not describe Elastic as an author or a partner.

The rc.96 fixes

The release calls itself a fix-and-honesty sweep since v1.0.0-rc.95. From the notes and the changelog:

• Inferred text fields take the Elasticsearch dynamic-mapping default doc_values: false (pull request 1329). term and terms on text with doc_values: true no longer fall to zero hits after a flush (pull request 1489).
• Tie order on two more term routes keeps score, then sequence number, then id (pull request 1393).
• autoindex adds a finalize read-back progress phase and warns, without aborting, when the server looks older than the client. The changelog says the walk’s cost is still open.
• On Windows, home falls back to %USERPROFILE% when HOME is unset (pull request 1427).
• xerj init backs up .mcp.json and AGENTS.md first, and leaves them untouched if the backup fails (pull request 1411).
• The security section bumps quinn-proto from 0.11.14 to 0.11.19, and calamine from 0.26.1 to 0.36.1 with zip from 2 to 8.

The same notes record a failed self-check: a wrong-and-confident detection gate scored 0.2449 against a ceiling of 0.05. The release says the watcher mechanism is real, and that this score is not enough to ship a detection watch as a confidently correct default. xerj.org says the built-in embedder is lexical and hash-based. llms.txt says --embed-mode neural is how a node leaves that default.

Corpus Hub is off in this build

The changelog says Corpus Hub v2 is not part of this release. Pull request 1423 is in the tree, and the build has no hub signing key, so no pre-indexed pack is verified or downloaded. Pull request 1491 makes xerj ask, xerj plus a question in quotes, xerj hub sync, and xerj hub pull exit 5 with one plain message. The MCP hub tools return that message. xerj init no longer writes the hub skill, and xerj --help says the command is unavailable. The changelog says xerj code and xerj corpus still work as before.

llms.txt, opened the same afternoon, still describes a hub path. For this tag, the changelog is the one that matches the binary: those hub commands are unavailable.

The token numbers are XERJ’s

The repository description says XERJ uses “40x fewer tokens than grep.” The case study prints a smaller comparison, and it is still XERJ’s measurement. It was not rerun here.

That page reports eight tasks in Python, JavaScript, C, and Java, 16 runs per arm, counting output tokens from claude -p. It shows 26,477 output tokens when Claude Code greps the source and 9,982 with a retrieved passage, called 2.7 times fewer, with 16 of 16 tasks solved on both arms. Memory alone solved 11 of 16. On the library it calls sift, memory solved 0 of 9 and retrieval solved 9 of 9. llms.txt cites the 2.7 times figure and “16/16 tasks solved.”

A second table on the same page covers 21 unfamiliar-API runs: memory at $21.90 solved 1 of 21, retrieval at $3.38 solved 21 of 21. Native Claude Code, grepping, also solved 21 of 21, at $4.26. The page’s “6.5 times cheaper” line is $21.90 against $3.38. The same-solve-rate comparison with grep is the 2.7 times figure.