Search in milliseconds on your own CPU, local and offline, MCP-ready. XERJ Query-1 turns questions into small programs and answers with the evidence.
Set up XERJ for me, following https://xerj.org/llms.txt. 1. Install XERJ with the XERJ Query-1 model and start it. 2. Index the folder I point you to (ask before reading anything outside it) and run one search. 3. Write three facts from that data into a facts namespace and ask XERJ Query-1 one question about them. Show me the answer, the program and the evidence. 4. Set up reference coding: index one open-source repo close to this project and look up one definition in it. 5. If this client takes MCP servers, register XERJ (llms.txt has the line for each client) and save one memory note through it. If you have no shell, use the HTTP or MCP steps in llms.txt instead.
Check it worked: curl -s localhost:9200/_facts | grep -o '"state":"loaded"' prints "state":"loaded" (empty for a few seconds while the model loads).
No shell: HTTP and MCP steps in llms.txt. Install it yourself: curl -fsSL https://xerj.org/get | sh -s -- --with-model on Linux and macOS; Windows and other targets: install.
Why now
Agents now do a growing share of the searching. Every answer they get from a language model costs seconds and money, and comes back as text nobody checked. XERJ moves that work into the index: search in milliseconds, and answers computed as small programs over your own facts, each with the evidence it used.
For developers whose coding agents need real code, and for teams whose data cannot leave their network.
XERJ Query-1 writes a short program and the engine runs it. No fact, no answer: it says UNKNOWN.
Search, vectors, facts, memory and an Elasticsearch-compatible API in one Rust binary.
Runs offline on a CPU, on a laptop or a server inside your network.
Measured
| What | XERJ | The usual way | How we got it |
|---|---|---|---|
| Time from question to answer | about 220 ms, XERJ Query-1 on 4 laptop CPU cores | 1.5 to 2.3 s, Opus 5.5 writing a RAG answer (API time) | Measured, same 160 questions |
| Cost per million questions | $42 of CPU, not counting writing the facts | $13,400 to $21,000 RAG; $65,400 GraphRAG; up to $694,000 with Opus reading the whole store | Priced from measured tokens and list prices |
| Wrong answers out of 160 | 1 (35 declined) | 9 to 15, RAG on Opus 5.5 searching one merged store. Opus 5.5 given the same records: 5; our sample can't tell 1 from 5 apart. | Measured on our development set; Query-1 was given each question's own records |
| What you run | 1 binary | 3 services: vector DB, graph DB, cache | Architecture, not a benchmark |
Data sets, method and caveats: the XERJ Query-1 post. Engine benchmarks, wins and losses alike: benchmarks.
Use cases
01 · Answers
XERJ Query-1 turns a question into a short program and the engine runs it over facts you or your agent wrote. No backing facts: UNKNOWN. Facts that disagree: CONFLICT.
final_*, program and evidence keys. Captured 2026-11-07.02 · Reference coding
Index the open-source repos closest to what you build, and on code it hasn't memorised your coding agent reads the real definition instead of inventing one and looping. Ready-made corpora are on the Corpus Hub.
2.7x fewer output tokens
than Claude Code grepping the same source, both 16 of 16 solved; 8 tasks in 4 languages on libraries written for the test. On code the model already knows, retrieval saved nothing.
xerj def on the rc.101 release over a clone of sled, captured 2026-11-07.03 · Search
xerj autoindex identifies each file by its content and indexes PDF, Office, HTML, CSV, JSON, email exports, SQLite, SQL dumps and code, from a folder or from S3, R2 or MinIO. Full-text and hybrid search.
4.6 s
to make 13 public NIST and CMS PDFs (22.9 MB, 3,032 records) searchable with the rc.101 release on a 32-core x86_64 laptop; then 0.9 ms at p50 over 200 different two-word searches. Run 2026-11-07.
04 · Live data Ships in rc.102
xerj connect postgres://... takes a consistent snapshot, then follows every insert, update and delete through Postgres logical replication, and picks up again after a restart.
19.7 ms
median from Postgres commit to readable in XERJ, 60 updates on a 200,000-row synthetic table, Postgres 16 (p95 21.7 ms; run 2026-11-05).
05 · Memory
Agents store, recall and forget notes, and link them into a graph where every edge keeps its evidence. Recall works by keyword and word overlap by default; turn on a neural encoder to recall by meaning.
24 MCP tools
on one node, next to search and facts (23 when XERJ Query-1 is off). Setup per client in llms.txt.
GET /_graph/agent1/ego and xerj mcp tools/list on the rc.101 release with the model on, captured 2026-11-07.06 · Enterprise on-prem Cohere hooks ship in rc.102
One XERJ node indexes your documents and databases on a CPU server you own, and passes the best hits to Cohere Rerank, Embed and Command, at a private endpoint, only when you turn them on. XERJ is not affiliated with Cohere.
787 MB of RAM
for the search node holding a live 200,000-row table, 3,033 document records and their vectors; the node running XERJ Query-1 used 1.05 GB (run 2026-11-05). Cohere calls were tested against a mock of Cohere's API.
07 · Bridge
XERJ answers the Elasticsearch 8.x REST API for indexing, search, aggregations and kNN, so existing clients and scripts can keep working while you move to the native API. It is its own engine, written from scratch in Rust.
0 failed
in our Elasticsearch 8.x conformance suite on main: 1,448 of 1,458 cases passed, 10 skipped (CI, 2026-11-07). No client version matrix yet, and the Python client's bulk helper sends PUT /_bulk, which rc.101 refuses; POST works (fix: #3036).
elasticsearch 8.19.3 Python client against the rc.101 release, captured 2026-11-07.What is open
Apache-2.0, on GitHub. Build it, audit it, fork it. Source
Ships as a separate download with each release.
Runs offline. Text goes out only to a model you turn on, at an endpoint you choose. Security policy
120
contributors
100
1.0 release candidates since 8 July 2026
14,239
engine binary downloads from GitHub releases
XERJ is building the search layer agents run on. If you want to deploy it, build on it or invest, write to us.