cerveau docs site github
Start here

Concepts

Cerveau is not a general harness with a local mode. It targets one situation — consumer GPUs, lots of system RAM, small mixture-of-experts models — and every layer is allowed to assume it.

The restrictions are the feature

Most agent harnesses are Swiss-army knives, and rightly so for their goal: they serve every provider, every model size, every deployment. But generality has a quiet cost. Machinery built to work with any model ends up implicitly designed around the strongest ones, and a small local model inherits scaffolding that assumes it will not fail — then fails in ways nothing catches.

Cerveau makes the opposite trade. Because the target is fixed, every layer gets to assume it:

Nothing is provider-agnostic, and that is the point. The result is not a smarter model — it is a harness that absorbs the failure modes of modest models, so a few billion active parameters deliver work that otherwise needs a much larger one.

Capability is engineered, not purchased

The industry's answer to a weak agent is a bigger model, a better card, or someone else's API. Cerveau's answer is engineering. Three examples, all of which have their own page:

Grammar-constrained memory
The distill pass runs against a compiled grammar, so prose is not a shape the model can produce. Memory
Compiled safety dispatch
The guard is Go code reading real arguments before anything reaches a shell. Safety guard
Pointer demotion
Bulky output leaves the window as a pointer to disk, rather than being summarised into something lossy. Context window

The whole machine, not just the GPU

The prevailing wisdom is that a model either fits in VRAM or it does not run. That is true of dense models. It is not true of a mixture of experts, where only a fraction of the parameters is active for any given token — and where the rest can live in system memory without the whole thing falling apart.

Your 64–128 GB of system RAM becomes model capacity. See Brain Cores for the measured trade.

Engineering over prompting

The distinction that runs through the whole codebase: a mechanism is something the model cannot argue its way out of; an instruction is something it can.

Instead ofCerveau does
"Please reply in JSON"A GBNF grammar on the sampler
"Do not run destructive commands"A guard at the dispatch point
"Do not use bash in this mode"Mode fencing in the registry
"Summarise what you are about to lose"A briefing assembled from the log
"Remember to search your memory"System-owned recall on every turn

Honest degradation

Every subsystem has a defined failure mode, and none of them is silence:

The reasoning is the same each time: a user who cannot see that something was lost cannot tell it apart from the model ignoring what it was told.

Local-first, and what that costs

One binary, bound to 127.0.0.1, no accounts, no telemetry, no outbound call. Unplug the network and it still runs.

The honest cost: the API is unauthenticated on loopback and can run shell commands, so access to that port is access to your shell. Reaching it from another device is a pairing problem with real cryptography, not a port forward.

What Cerveau is not