For two years, useful AI meant renting reasoning, memory, and tools from a few companies you could not afford to leave. That was defensible while open models trailed the frontier by a wide margin. The gap has nearly closed: capable models now run on a laptop. So the thing worth building is no longer a better model. It is the open infrastructure that lets a person own the capability, run it on hardware they hold, and keep it working after they unplug from the frontier labs.
The work you build into a rented model stays the vendor's, not yours. Caletta turns that around with a loop you own: routing sends each call to a frontier API or to a model on your own silicon; capture turns your usage into a training set you keep; training runs on hardware you control. Four beliefs decide what we ship and what we refuse to.
Capability you own, not capability you rent
The transport substrate for this loop is open and running today: mlx-go-iroh for peer-to-peer mesh transport over go-iroh, and localtinker as a local training coordinator that serves a training API on your own machine instead of a hosted service. The Go-native inference runtime for Apple silicon is built and in release preparation. We say which is which.
Compute no one else can switch off
Intelligence that runs on hardware people hold outlives any single cloud, company, or country deciding to cut access, the way a file shared across a network outlives any one machine. The transport for that is the published mesh layer. The first measured step across it exists: two machines trained and merged one adapter to identical bytes across three rounds, documented with its failures in Two Macs, one adapter. A demonstration that the datapath keeps making progress after a node drops is work still ahead, and we describe it as such rather than claiming it is done. Where this belief is pointed long-term is written out in Democratized distributed training.
Inspectable, never trust-me
A claim about a model should be checkable, not taken on faith. Our runs emit evidence: network rows, logs, traces, and reproducible receipts. We are careful about what that evidence proves. A receipt is reproducibility evidence, not hardware attestation, and it does not prove that no hidden step occurred. We would rather state the bound than oversell the guarantee. That distinction has its own note.
On your side of the line
Private by default is a boundary, not a setting. The security layer, policy, brokered secrets, routed egress, and operator-owned evidence, exists so the loop runs on your side of the trust boundary and stays there.
Everything above is Go, in the open, built to run with no vendor API in the loop. The longest-standing proof a reader can check is langchaingo, the Go LLM framework, alongside the published transport substrate. The pieces still in release preparation are named as such. Nothing here links to a repository you cannot clone today.