Recce vs Data Workers: Data-Change Review Compared

Recce is an open-source data-change review toolkit built around dbt pull requests, with a purpose-built review UI and a managed cloud. Data Workers' Data Guardrail Agent reviews data changes on any of four warehouses with plain SQL — no dbt required — and ends every review with a machine-actionable verdict. Both catch silent data damage before it merges.

The two tools attack the same failure mode — a PR that passes code review while quietly changing the data underneath it — from genuinely different angles. Recce optimizes for a rich human review experience inside the dbt workflow. The Data Guardrail Agent optimizes for warehouse-agnostic evidence gathering and structured handoffs to other agents. This comparison lays out both honestly, because the right choice depends on your stack and your review culture, not on either vendor's marketing.

What Recce is

Recce is an Apache-2.0 open-source toolkit (github.com/DataRecce/recce) for reviewing the data impact of dbt pull requests. It was founded by CL Kao — the creator of the SVK version-control system, who previously built PipeRider at InfuseAI — and the company raised $4M in April 2025 in a round led by Heavybit with participation from Vertex Ventures US and Hive Ventures. The premise is that dbt PRs deserve the same review discipline as application code, and that reviewing them requires comparing actual data between your base and PR environments, not just reading SQL.

Recce's check taxonomy is the most mature part of the product. Out of the box it runs lineage diffs, schema diffs, row count diffs, profile diffs, value diffs, top-k diffs, histogram diffs, and arbitrary query diffs. Checks accumulate into a checklist that documents the review, and preset checks defined in recce.yml standardize what every PR gets examined for. The recce server command launches a local review UI with column-level lineage, so a reviewer can see exactly which downstream models a change touches before approving it.

Recce Cloud is the managed layer: it runs AI data reviews on pull requests and posts the results as CI bot comments. As of June 2026, Recce's pricing lists a Free tier at $0 with 100 agent reviews per month, a Team tier at $250/month billed annually with 1,000 reviews per month, and a custom-priced Enterprise tier that adds SSO, BYOC deployment, and RBAC. Recce also ships a Claude Code plugin and an MCP server, so agentic coding workflows can drive reviews directly. The company claims SOC 2 compliance.

Recce's main constraint is that it is built around dbt artifacts. If your transformations live in dbt, that is a feature — the tool understands your manifest, your models, and your environments natively. If they do not, Recce is largely off the table today; SQLMesh support exists but is experimental.

What the Data Guardrail Agent is

The Data Guardrail Agent is the data-change review agent in the Data Workers swarm. It runs reviews on Snowflake, BigQuery, PostgreSQL, and Databricks using plain SQL and INFORMATION_SCHEMA — there is no dbt requirement. If dbt manifests are present, they enrich the review with model-level context, but the agent works the same against a pair of schemas, a list of tables, or a raw diff.

Its defining design choice is the progressive, cost-aware review. The agent starts with evidence that costs nothing — metadata from the information schema — and only descends to row counts, then column profiles, then row-level proof when the cheaper layer cannot settle the verdict. Follow-up queries are capped, so a review never turns into an unbounded warehouse bill. Every check is re-runnable, and every verdict is stamped with the warehouse mode it actually ran in, so you always know whether a conclusion came from live warehouse evidence or from metadata alone.

• •Risk buckets, not vibes — every finding lands in HIGH, MEDIUM, or LOW, governed by the limited-evidence rule: missing evidence widens uncertainty, it never shrinks risk. A review that could not gather row-level proof cannot quietly downgrade a HIGH.
• •A handoff plan, not a comment — every verdict carries machine-actionable structured tool calls addressed to the swarm's schema, incident, quality, and governance agents. When the change is safe, the plan is empty. When it is not, remediation can start without a human re-typing the findings into another tool.
• •MCP-native — it ships as an MCP stdio agent, so it runs inside Claude Code, Cursor, or a CI job with the same interface.

Honesty requires stating the gaps too. The Data Guardrail Agent has no dedicated review UI today — verdicts render as PR-comment markdown plus structured JSON, which is great for machines and adequate for humans, but it is not Recce's purpose-built review surface. There is no managed cloud review service. And dbt selector support is partial: the agent consumes manifests, but it does not yet speak the full selector syntax a deep dbt shop may expect. It is also not open source.

Recce vs Data Guardrail Agent: side by side

When Recce is the better fit

Choose Recce when your transformation layer is dbt and your bottleneck is the human review experience. The review UI is genuinely good: column-level lineage, a mature taxonomy of eight check types, and a checklist that turns an ad-hoc review into a documented artifact. Preset checks in recce.yml give a team a shared definition of 'reviewed,' which matters more than any single diff. If you want a managed service that comments on PRs without you running anything, Recce Cloud's free tier is a low-friction starting point, and the Apache-2.0 core means you can adopt the toolkit without a procurement conversation.

Recce is also the safer choice if you want an open-source tool you can read, fork, and extend — the Data Guardrail Agent is not open source. And teams that already collaborate in a review-UI culture (think GitHub's PR view, but for data) will find Recce's workflow familiar on day one.

When the Data Guardrail Agent is the better fit

Choose the Data Guardrail Agent when dbt is not your whole world. Plenty of real data estates — stored procedures, Airflow-orchestrated SQL, Spark jobs writing to Databricks, vendor pipelines landing in BigQuery — never touch a dbt manifest, and the agent reviews changes to all of them with the same plain-SQL approach. It is also the better fit when warehouse cost discipline matters: the progressive review spends nothing before it spends something, and capped follow-ups make the cost of a review predictable.

The deeper differentiator is what happens after a finding. Recce ends a review with information for a human. The Data Guardrail Agent ends a review with a handoff plan — structured tool calls that the rest of the Data Workers swarm can execute. If a HIGH-risk schema change needs an incident opened, a quality check added, and a governance review, those are tool calls in the verdict, not action items in a comment. For teams building toward autonomous data operations rather than better-informed manual ones, that is the point of the product.

Finally, if your reviews happen inside agentic coding tools — Claude Code, Cursor, CI — the MCP stdio interface means the guardrail runs where the change is being made, with re-runnable checks and honest warehouse-mode stamping so nobody mistakes a metadata-only pass for row-level proof.

Is Recce open source?

Yes. The core Recce toolkit is Apache-2.0 licensed and developed in the open at github.com/DataRecce/recce. Recce Cloud — the managed AI-review service with PR bot comments — is the commercial layer on top. The Data Guardrail Agent, by contrast, is not open source; it ships as part of the Data Workers swarm.

Does the Data Guardrail Agent require dbt?

No. It runs data-change reviews on Snowflake, BigQuery, PostgreSQL, and Databricks using plain SQL and INFORMATION_SCHEMA. When dbt manifests are available they enrich the review — model context, declared changes versus observed changes — but they are an input, not a prerequisite. Recce takes the opposite stance: it is built around dbt artifacts, with experimental SQLMesh support.

Can you use both together?

Reasonably, yes. A dbt-centric team could use Recce's review UI and checklist for human review of dbt PRs, and run the Data Guardrail Agent in the same CI pipeline for non-dbt changes — direct DDL, ingestion schema drift, warehouse-side migrations — or wherever a structured, machine-actionable verdict is needed for downstream automation. Both expose MCP servers, so both compose with agentic workflows rather than competing for the same seat. Our resources library covers how review fits into the broader autonomous data-engineering loop.

Recce is the stronger choice for dbt teams who want a mature, open-source review UI and a managed PR bot. The Data Guardrail Agent is the stronger choice for mixed-stack teams who want cost-capped reviews on four warehouses without a dbt dependency, and verdicts that other agents can act on. Both beat the status quo, which is approving data changes by reading SQL and hoping.