A.R.E.S · AGENT RUNTIME · OPEN SOURCE

Agents run here.

One interface across many model providers, with tool calling, retrieval, MCP and per-tenant metering, in Rust, light enough to run on a small VPS. Agents are defined in configuration and read from Eruka before every generation.

View on GitHub Talk to us

MITON CRATES.IOIN PRODUCTIONMULTI-TENANT METERING

WHAT IT DOES

One runtime under the agents we ship.

Providers and portability

One interface across Anthropic, OpenAI, Google, NVIDIA NIM, Groq, Bedrock and local models, so an agent can move when prices or terms change.

Tools, retrieval and MCP

Tool calling, retrieval, MCP client and server, structured outputs and streaming.

Multi-tenant metering

Every client isolated, every run measured, every cost attributed.

About fifty megabytes at idle, sub-second cold start, runs on a small VPS.

GOVERNED BEFORE GENERATION

Reads from Eruka before every generation.

Developers get a runtime they can read, self-host and extend, with Eruka as the governed store behind it.

REQUEST
RUNTIME
GOVERNED STORE
OUTPUT

How the write gate works

CONFIGURATION, NOT CODE

Declared in one TOML file.

Providers, the models on them, and the agents that use those models. This is the shape of the file, not a working one.

toml · outline

[providers.local]   # a named provider: its type, address and default model
[models.balanced]   # a named model on that provider
[agents.tutor]      # an agent: its model, its tools, its system prompt

The full example is in the repository↗

WHAT IS SOLD

The runtime is free. Operating it is the service.

Run it yourself: FREE · MIT; The runtime is free to run yourself.
Inside our services: SOLD NOW; It runs as the managed runtime under every client system, in-region on our infrastructure or in your own cloud account.
Managed ARES for agencies: DESIGNED; White-label, for agencies running agents for their own clients.