Correctness over throughput.
Whatever melt returns equals what Snowflake would have returned. A parity sampler dual-runs a fraction of routed queries and alerts on drift. We refuse to federate anything that touches policy-protected tables.
About melt
The unified Snowflake cost layer.
Melt is an open-source proxy that sits in front of Snowflake and decides, per query, where it should run — the cheap lakehouse, the right-sized warehouse, or both. Drop-in for any Snowflake driver. Transparent to dbt, BI tools, and agents.
Why nowAgents broke the assumptions
Agents broke the assumptions
Snowflake billing was built on.
For a decade, warehouse demand had natural throttles. Analysts refreshed dashboards a few times a day. dbt models materialised overnight. Ad-hoc queries fired when someone asked a question.
Autonomous agents don't run on that cadence — they run per prompt. A single conversation can fan out into dozens of small filters, joins, and aggregates. Across a fleet of agents, the warehouse never gets to spin down.
Most of those queries don't actually need Snowflake compute. The bet behind melt is that a routing layer between the driver and Snowflake — invisible to the agent issuing the query — can save the bill without changing a connection string.
What we believe
Three principles, no compromise.
Open by default.
The proxy, the sync, and the CLI are Apache-2.0. Self-hosted by default. No phone-home telemetry. The only data that leaves your environment is what you choose to send to the hosted control plane.
Transparent to your stack.
Drop-in for the official Snowflake driver, JDBC, ODBC, dbt, Looker, Sigma, Hex. Your connection string changes; nothing else does. No SQL rewriting, no app-side routing logic, no warehouse downtime.
Who's building it
A small team, an open repo.
We're a tiny crew of data-infra engineers shipping in public, with a growing community of contributors and design partners running melt in production.
Contributors & design partnersgrowing · open repo
Melt ships in public. Engineers from data teams running it in production file issues, PRs, and field reports against the open repo every week — that's where most of melt's sharpest features come from.
Contribute on GitHubBuild with us, or just use it.
Melt is open-source and self-hosted. Pull the binary, point your driver at it, and watch the credits drop. Or talk to us if you want it run for you.