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.
02Governed
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.
03Efficient
Lower AI spend by construction.Reruns hit cache — Snowflake doesn’t bill the same answer twice — and a smaller model does the work: 60% fewer tokens on DABStep.
01Local-first by design
No new vendor in your data path.
A native Mac app on your own LLM subscription and warehouse credentials — no vendor cloud to clear through security review. Git-native: catalogs are repos, promotion to main is the governance event. Bulk data stays in the warehouse; only aggregates cross into the local sandbox.
xorq_snow
11 varson this Mac
TASKSCATALOGSPROFILES
YOUR KEYS 11
ANTHROPIC_API_KEYsk-ant-••••••••••
SNOWFLAKE_ACCOUNT••••••••••••
SNOWFLAKE_PRIVATE_KEY••••••••••••••••
DATA RESIDENCY
Warehousebulk rows stay here
aggregates →
Local sandboxcompute on your Mac
no proxy · no token markup
02The 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 at lower model spend.
DABStep — 450 questions over payment data. Same model, same prompt, only the catalog changes — and Haiku + catalog beats Sonnet baseline on 60% fewer tokens and half the turns. Accuracy and cost savings come from the same place: the catalog.
Governed AI usually means engineers standing up MCP servers, permission services, and agent rigs. With Xorq that layer comes with the install: every tool call routes through the semantic catalog and clears verification gates before an answer reaches your team. The LLM is swappable; the harness is what we ship.
whats the heaviest penguinfinished
ORCHESTRATOR
The heaviest penguin is a male Gentoo weighing 6,300 g — Biscoe island, recorded 2007.
DATA QUALITY
QUALITY OK
SOURCE-PENGUINS n=344 · nulls body_mass_g=2 · range 2700–6300g valid · no impossible values
VERIFICATION
VERIFIED
Confirmed verbatim from VERIFY_SOURCE-PENGUINS_HEAVIEST
orchestrator verifier quality-checker
04Cited by construction
Unbreakable lineage.
Agents compose only on sources a human blessed. Every answer ties back to a cataloged expression — source, cache, schema — so when an auditor asks where a number came from, the answer is one click, reproducible six months later.
composed51241c900965flights
DATALINEAGE
THIS ENTRY
51241c900965 · composed · 2 cols
COMPOSED FROM 1
SOURCE-FLIGHTS
CACHE
letsql_cache-snapshot-d22bf8d…0dfa7b · parquet/
SCHEMA 2
airportstringtotal_flightsint64
OPERATION GRAPH 2
engineread
engine RemoteTable 16 colssource: @9f425664259a
read Read 16 cols
05Replay any answer
Spans, cache hits, replayable.
Open any answer for the spans, hits, and durations behind it. Replay a single tool call. Diff against last week.
replayedwhats the busiest airport
RESULTTRACE
Spans15
Cache hits1
Hit rate100%
Span · waterfallDuration
catalog959ms
to_pyarrow684ms
_register_node659ms
cache_hit659ms
compile<1ms
06The open-source core · Apache-2.0
An executable memory layer for data work.
Markdown memory burns tokens: the agent reads its notes, writes a one-off script, runs it, and throws it away. Xorq’s memory is executable — durable, content-addressed expressions that run as-is, so the next agent hits cache instead of re-deriving. That’s where the tokens and turns go.
Declarative Ibis expressions that run the same on your laptop or your warehouse. The semantic catalog that powers Xorq Desktop.
× XORQ CONSOLE / INBOX
▸ How can I help with Inbox?
HAIKU ▾↑
2 CATALOGS ▾
16 TOOLS ▾
9 SKILLS ▾
3 QA GATES ▾
CACHED RESULTS17 ENTRIES / 341 KB
›[LETSQL_CACHE]d86db3141 KB11H AGO
›[SNAPSHOT]6544f7b612 KB12H AGO
›[SNAPSHOT]af2021a34 KB12H AGO
›[SNAPSHOT]76e92a4b4 KB12H AGO
›[SNAPSHOT]6ff65fb96 KB12H AGO
07Things people ask first
What is Xorq Desktop?
A native Mac binary. Point it at your data, pick a workflow, and an agent works against a shared catalog of cited 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. The catalog, gates, caching, and warehouse connections come with the install — nothing to build or babysit.
How does this get through security review?
It runs on the analyst’s Mac with your existing warehouse credentials and LLM subscription — no vendor cloud in the data path, so no new data processor to assess.
What is the “semantic catalog”?
An immutable store of executable semantics: expression + metadata + cached result, addressed by hash. Metric definitions live here as runnable code; agents compose new entries on top of old ones.
How is this different from a notebook + an LLM?
Notebooks accumulate output. Xorq accumulates work. Last week’s expression is still here, still addressable; the next agent picks up where the last one left off.
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.
08Two paths to the same harness
Pick how you adopt.
Run our integrated harness, or layer xorq’s primitives into the harness you already use. Same catalog. Same open-source library.
01Our harnessprivate 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.
$ open Xorq.app
02BYO harnessopen source
Claude Code × xorq
Already in Claude Code? Layer the same catalog, lineage, and guardrails in as a plugin. No app, no migration.
