Intelligence needs
an environment.
Atreyu connects models, agents, tools, and memory into a platform where conversations can become ongoing work.
Beyond the chat window.
A model response is one part of the story. Atreyu also manages the work it starts, the context it carries, and the authority it operates under.
An environment for AI that can converse, act, and carry work forward.
↗02Architecture atlasFour interacting planes. Clear owners for state and execution.
↗03Workflows & toolsStable contracts between a model’s request and an operation’s effects.
↗04Context & continuityHistory, memory, and derived context each have their own job.
↗05Agents & governanceContinuity and agency need explicit authority boundaries.
↗06Access AtreyuChoose an environment, then use its own account and access flow.
↗More life from every
layer of memory.
Our ROCm-optimized llama.cpp fork powers Atreyu’s machinazero:primary model/provider. Built around AMD Instinct MI50 / gfx906, it combines three-tier prompt-state caching with targeted inference optimizations.
Keep active work close.
Live inference state on the GPUs.
Retain useful context.
Reusable range checkpoints in system memory.
Bring cold context back.
Asynchronous backing for completed checkpoint writes.
One project.
Distinct environments.
This is Atreyu’s public home for information and documentation. The applications run separately.
Explore the current platform
The actively developed Atreyu environment. Features evolve here; use its own account and access flow.
Open llm.viper-7.com ↗Returning to an existing setup?
The older deployment, maintained less actively. It may differ from development and these documents.
Open api.netbubbles.com ↗