Colrows puts a compiler—and a billable question—between agents and SQL

Colrows has surfaced as a new enterprise text-to-SQL entrant with a distinct pitch: natural-language questions should be compiled against a versioned semantic graph before SQL reaches the warehouse. Its August v1.702 release makes Model Context Protocol (MCP) the supported agent-integration surface and adds document processing plus a shared analytics cache. The product appeared on Hacker News on August 28, but the more useful practitioner evidence is in Colrows’ own changelog, documentation and pricing terms.

What shipped

The v1.702 MCP server exposes REST endpoints, OAuth discovery, telemetry, cursor-based continuation and tools for metadata inspection, column profiles, join discovery and governed data access. Colrows says it retired cluster-based CLI clients in favor of MCP. The same release renames analyze_data to ask_data, changes join-discovery response models and stops creating a database automatically through the organization CLI. Existing integrations therefore need client regeneration or adapter changes—not just a new endpoint.

Underneath that interface, Colrows documents a four-stage path: intent parsing, semantic resolution, constrained planning and governed execution. RBAC, ABAC and row- or column-level predicates shape the allowed graph before SQL is generated; ambiguous join paths are intended to fail compilation rather than silently choose one. The platform lists Snowflake, Databricks, BigQuery, Postgres, ClickHouse, Trino, Oracle and SQL Server among its backends.

The commercial boundary matters

Colrows’ SaaS pricing page meters accepted natural-language requests rather than tokens, seats or dashboard views. Re-running an existing resolution in a dashboard is unmetered, while repeated questions and agent calls count again. Published dollar prices are not listed. The company says service continues through 110% of monthly capacity, after which new questions pause unless the customer approves more capacity; existing dashboards continue.

That makes request classification part of architecture. Teams should decide whether retries, clarification turns, scheduled agents and cache misses create new metered requests before putting Colrows behind a high-volume agent. They should also regression-test OAuth scopes, continuation state and renamed tools during the v1.702 upgrade.

Treat the accuracy numbers as vendor evidence

Colrows reports 98.2% execution accuracy for its compiler versus 14.5% for raw LLM text-to-SQL across 50 proprietary schemas plus MIMIC-III and eICU data, and 100% join-path accuracy for the compiled route. Those are first-party results, not an independent leaderboard submission. The company publishes the test framing, but buyers still need to reproduce it on their own schemas, policies and dialects.

The launch is notable less for another chat interface than for where it draws the control plane: below the prompt, before SQL generation, with a semantic graph and policy compiler shared by agents and dashboards. The evaluation question is consequently concrete: does the compiler reject ambiguity and preserve authorization under real schema drift—and can the request meter stay predictable when agents retry?