Waitlist · Desktop app preview

AI generates numbers. Xorq verifies them.

you · dig for it

which script? · no lineage

xorq · verifies it

Open source · github.com/xorq-labs/xorq

Snowflake Databricks DuckDB Postgres 6+ more

01 Verified

A verifier checks every number. Deterministic checks for data quality, lineage, and reproducibility run alongside the work, so a hallucinated total gets flagged and corrected before anyone acts on it.

02 Governed

Built only on blessed sources. Agents compose new expressions only on sources a human signed off on, so ground truth stays under change control with an audit trail by construction.

03 Efficient

Cached runs. Fewer tokens. Every result is cached and reused, so the warehouse never bills the same answer twice. Separately, the semantic catalog helps a smaller model do the work: 60% fewer tokens on DABStep.

01 Xorq Desktop · local first

No vendor cloud required in your data path.

A native desktop app that uses your own LLM subscription and warehouse credentials, reducing the scope of security review. Built on Git: the catalog is a repo, and promotion to main is human-controlled. Bulk data stays in the warehouse; only required aggregates enter your customer-controlled local sandbox.

xorq_snow 11 vars on this machine

TASKS CATALOGS PROFILES

YOUR KEYS 11

LLM_API_KEYsk-•••••••••••••

SNOWFLAKE_ACCOUNT••••••••••••

SNOWFLAKE_PRIVATE_KEY••••••••••••••••

DATA RESIDENCY

Warehouse bulk rows stay here

aggregates →

Local sandbox compute on your machine

no proxy · no token markup

02 Xorq Memory · semantic catalog · Apache 2.0

Xorq Memory == Semantic Catalog

Xorq Memory is an executable semantic catalog. Every expression stores its schema, SQL, sources, are content-addressed, so each number traces back to the work that produced it.

Unlike static documentation, this memory runs on your laptop, or wherever your agents run with headless offering. Agents reuse governed expressions and cached results instead of starting over.

Built on Ibis. Powers Xorq Desktop.

Star —

pennybank @ ac9a620 shared

filter alias / hash / kind / metadata… 10

10 entries · 10 aliases · 20 sql · 29 src · 10 cached

kindaliasmetahash

source transactions 2 src1 cache 7be9896e8edd

source merchants 2 src1 cache aefa6363c5f8

source accounts 2 src1 cache b71cb6d441db

expr_builder semantic-payments 1 bld1 cache efb05bc7b6f3

expr revenue-by-merchant 5 src1 cache 5c645bf91ab9

expr chargebacks-by-quarter 3 src1 cache df8931673a34

append-only · content-addressed · nothing overwritten

03 Unbreakable lineage

Lineage across every hop.

Agents compose only on sources a human blessed. Every answer ties back to an entry in Xorq Catalog, with its source, cache, and schema, so when an auditor asks where a number came from, the evidence is one click away and can be reproduced from versioned inputs.

Open any answer for the spans, hits, and durations behind it. Replay a single tool call. Diff against last week. Because verified work is cached by design, a repeated question replays instead of rerunning.

No rework, fewer tokens spent in reconstructing scripts.

semantic-payments expr_builder pennybank

METADATA LINEAGE

6 relations · 5 links · 46 cols · 1 source

source Read

reads the blessed source · 46 cols

source: @43057a62acdb semantic model

op CachedNode

reads a cached result · 46 cols

op FlightUDXF

streams over Arrow Flight · 46 cols

op Project

projects relation columns · 1 col

leaf RemoteTable this entry

runs on the query engine · 1 col

04 The proof, in benchmarks

DABStep @dabstep

450 questions over payment data · accuracy %

Haiku baseline50%

Sonnet baseline75%

Haiku + Xorq catalog84%

same model · same prompt · only the catalog changes

Better answers without a bigger model.

DABStep: 450 questions over payment data. Haiku + Xorq catalog scores 84%, beating the 75% Sonnet baseline. Same model, same prompt, only the catalog changes.

Read the full story →

05 Xorq Headless · enterprise

The desktop is optional. The verification is not.

Verification harness in the the desktop does also runs headless in your cloud as subagent exposed as MCP server; a verification subagent computes the answer against the same catalog and hands back a certificate with the proof, lineage, and cache references behind it.

It drops into the stack you already run: your warehouse, model endpoint, Git server, and auth. No UI is required.

xorq_headless in your VPC

CLAIM CERTIFICATE LINEAGE

YOUR AGENT CLAIMS “2023Q1 revenue comes to $23.8M”

certificate · @revenue VERIFIED ✓

$23.8M

scalar · currency · tolerance ±$0.05M

witness on quarterly_revenue compose source.filter(_.quarter == '2023Q1') predicate select revenue where quarter == '2023Q1' cache HIT · warehouse not billed twice

recomputed from a blessed source · one turn · no UI in the loop

06 Deploy how you want

Pick how you adopt.

Run our integrated harness, layer xorq’s primitives into the harness you already use, or deploy it headless in your cloud. Same catalog. Same open source library.

01 Our harness · recommended private preview

Xorq Desktop

The integrated harness. Open the app, point it at your warehouse, ask. Agent, catalog, and lineage in one native binary. No terminal required.

Get started

$open Xorq.app

02 Plugin example open source

Claude Code × xorq

Use the Claude Code plugin today, or connect Xorq’s model-neutral catalog from Codex, Cursor, and other agent harnesses. No data migration.

>/plugin marketplace add xorq-labs/claude-plugins>/plugin install xorq@xorq-plugins

03 Headless enterprise

Xorq Headless

The same verification subagent, running as a service in your cloud. Your agents call it and verified results return with a certificate and lineage, on your warehouse, keys, and auth. Deployed with our team.

Things people ask first

What is Xorq Desktop?

A native desktop app. Point it at your data, pick a workflow, and an agent works against Xorq Catalog, a shared store of governed, executable expressions.

When can I use it?

Private preview. Waitlist for the desktop; the library (uv add xorq) ships today.

Do I need to migrate my data?

No. Connects to Snowflake, Databricks, DuckDB, S3, Postgres — whatever you already run.

Do I need to set up an MCP server?

No. Desktop packages the catalog, verification gates, caching, and warehouse connections; you configure it with your existing credentials.

How does this get through security review?

It runs on the analyst’s machine with your existing warehouse credentials and LLM subscription. Bulk data stays out of a new vendor cloud, reducing the scope of security review.

What is a semantic catalog?

A semantic catalog: a store of executable semantics, expression metadata, lineage, and cached results, addressed by hash. Metric definitions live here as runnable code; agents compose new entries on verified work.

How is this different from a notebook + an LLM?

Notebooks accumulate output. Xorq accumulates reusable work. Last week’s expression remains addressable, so the next agent can build on it instead of starting over.

Two people, same question — conflicts?

Same expression → same hash → automatic dedup. Identity is content, not file path.

Linux / Windows?

Desktop is Mac-first. The library and Textual TUI run everywhere Python does.

Verifiable AI for Enterprise Data Teams