Mid-market companies already own their data. It sits in an ERP, in SQL Server, in the warehouse they operate today. What they don’t own is the loop that turns a question into a report: a new question lands, a ticket opens, a consultant queues it, an extract is rebuilt and republished — and two to three weeks later the answer arrives, frozen, already slightly wrong.

We evaluated an industry-leading reporting platform in exactly this position: capable, widely deployed, and structurally dependent on a publish-an-extract cycle. Its published data sources actually stripped fields that existed in the customer’s own database — users were getting less analytical surface than their own ERP contained, not because the data wasn’t there, but because the publishing step omitted the joins.

The engagement question: could an AI-native engine read the same sources directly — governed, audited, read-only — and replace the authoring loop itself?

Before writing a single report, Claude agents documented the estate: 2,498 assets across the ERP database — tables, views, triggers, relationships — compiled into a navigable atlas an agent can traverse without ever guessing a column name. The incumbent platform’s report suite was reverse-engineered the same way: 716 calculated fields, 27 dashboards, 127 component views, mapped to their true SQL equivalents.

Then the rebuild: 24 of 24 production dashboards re-implemented as live, self-contained HTML surfaces in roughly three days — running on direct, read-only SQL against the source database. No extracts, no republish cycle, no copy of the data to keep in sync. Parity was checked against the incumbent platform’s own published numbers.

The rebuild surfaced findings the incumbent’s dashboards had never shown: eight figures of revenue accumulating under a placeholder salesperson record, a budget line frozen at the same value for twelve consecutive years, and dozens of deals marked closed-won after their expected close date — a “deals we save” pattern invisible in the extract layer.

The product of the engagement is GRT — our reporting engine, a plugin-hosting MCP server — and it is built as two tiers with a hard line between them.

The data plane is LLM-free. Every query is validated to a single read-only statement before it reaches the database; writes and DDL throw before execution. Connections run under least-privilege logins, scope-gated, tenant-isolated, fully audited. 282 guard scripts enforce this contract on every build.

Claude sits above it. The engine publishes MCP tools, prompts, and a progressive schema resource, so the caller’s own Claude — in Claude Code, Claude Desktop, or our web client — discovers real tables and joins before composing a query. It never guesses a column name, and the server never ships row data to a model it doesn’t control. Ask for a report split by branch instead of region, and it’s rewritten from plain English in seconds — inside the same governed rails.

In production topology, GRT registers as a plugin behind Cerberus, our orchestration platform: signed artifact closures, behavioral contract fingerprints, and an upstream identity assertion model, so the orchestrator can verify what a plugin is before trusting what it says.

GRT and Cerberus are owned by separate Claude Code agents, each with its own repository. The integration between them was negotiated the way senior engineers do it: versioned design memos, committed on both sides.

The correspondence reads like a good engineering org. The Cerberus agent cross-checked the GRT agent’s handoff against GRT’s source and recorded that the contract resource was “live and described faithfully — not aspirational.” The agents disagreed about where shared types should live, and one conceded on the merits, citing the architecture decision record that forbade the coupling. A reciprocal clean-room audit caught a latent cryptographic bug — a signature-verification path that would have fail-closed rejected valid artifacts after a future dependency upgrade — which was fixed and pinned with a regression guard. And where a decision belonged to the business, the agents stopped and flagged it for a human.

The build itself is measured, not vibes: 755 of the engine’s 1,005 commits carry a Claude co-author trailer, and a single metered multi-agent workflow — 33 agents, 950 model calls across Claude’s model family — cost $33.02 at published API rates, with prompt caching serving 40.5 million cached tokens against 2.7 million written.