Sign up now for early access to the Xorq Desktop app Xorq Desktop early access ./waitlist
XorqXorq
./waitlist

Agent Verification

A subagent that answers with receipts. Any agent asks; Xorq recomputes the claim against your data and answers with proof.
solutionsagent verification

A subagent that answers with receipts.

Your existing assistants and agent frameworks call Xorq to get lineage-backed answers from your data, so any agent's output is checkable at the source.

your agent → xorq verify mcp · one call
claim · from your agent "2023Q1 revenue comes to $23.8M"
certificate · @revenue VERIFIED ✓
$23.8M scalar · currency · tolerance ±$0.05M
witnessquarterly_revenue · blessed source predicateselect revenue where quarter == '2023Q1' cacheHIT · warehouse not billed twice
recomputed from a blessed source · one turn · no UI in the loop
$ xorq verify --claim

Your agent said $23.8M. Confidently. Now what?

today
01An agent reads a few tables and states a number in prose.
02Someone asks it to double-check. It restates the same number, or a different one.
03A second model is asked to grade the first. Now there are two opinions.
04An analyst opens a notebook and re-derives it by hand. Forty minutes.
no source not reproducible opinion, not proof
vs
with xorq
01Your agent hands the claim to Xorq in one MCP call.
02Xorq recomputes it from a blessed source on your warehouse.
03A certificate comes back: value, tolerance, lineage, cache reference.
04The agent continues with a number anyone can click to the source.
recomputed certified one turn
> claim → certificate

Any agent asks. Xorq answers with proof.

The subagent runs as an MCP server next to your agents, on your warehouse. It never asks the model to trust itself.

01 · claim Agent states a claim Any harness, any model: Claude, Codex, Cursor, or your own. It passes the claim and its context to Xorq.
02 · recompute Xorq runs the real query The claim is compiled to an expression against a blessed source and executed on your warehouse.
03 · certify Certificate returned Value, tolerance, lineage, and cache reference come back as a signed artifact in the Ledger.
04 · continue Agent proceeds, or a human decides Your agent proceeds with the verified value. Flagged claims route to a person before they ship.
# architecture · mcp subagent

Xorq sits beside your agents as a subagent. Nothing else in your stack moves.

Your harness, model endpoint, Git server, and auth stay where they are. Xorq is a remote MCP server the agent calls as a subagent: it sends a claim, the harness works it, and a structured response comes back over the same connection.

your agents
Claude Codexorq plugin
Codex · Cursormcp client
Your own frameworkLangGraph · CrewAI · custom
claim →remote MCPsend · next · close
xorq · sub-agent harnessMCP server
Memory · Ledgersemantic catalog · curated expressions · approvals
Query loopskills · tools · hooks · compiles the claim to a real query on any engine
Verifierrecomputes · checks tolerance · resolves lineage
Cacherepeated claims replay, not rerun
Gate · flagged claims route to a personhuman
← responsevia MCP
structured response
Certificatevalue · tolerance · verified ✓
Lineageevery hop to a blessed source
Prose + artifactsanswer text · Ledger entry · reusable by the next agent
real queries ↓ ↑ rowsyour warehouse · your LLM provider · your Git server (catalogs)
verifiesScalars, tables, trends, and comparisons an agent asserts from your data
runs onSnowflake, Databricks, BigQuery, DuckDB, Postgres, Trino
works withClaude Code, Codex, Cursor, and any MCP client; headless in your VPC
// ready when you are

Keep your agents. Add the proof.

Talk to us
other solutions:verified reportingdecision agents