Start

Introduction

Rakuvo is a gateway that sits between your AI agent and its model and makes the agent more reliable, without changing the agent's code.

#What it is

Your agent talks to a model through an OpenAI-compatible API. Rakuvo sits in the middle of that conversation. Every request passes through a pipeline of layers, each aimed at one way agents fail in practice: skipped tool calls, malformed arguments, runaway loops, overflowing context, forgetful sessions.

You change one line in your agent, its base URL, and nothing else.

#The six layers

LayerWhat it does
PromptAdds a short instruction so models use their tools instead of guessing.
ToolsChecks and repairs every tool call, and can ask a small judge model to veto one that makes no sense.
ContextKeeps long runs inside the model's window by masking old tool results.
Loop guardNotices an agent repeating itself and nudges, then stops it.
Model policyChooses which model answers a call. Not yet shown to work live.
MemoryRemembers what each user has said across sessions, as a graph of dated facts.

Each layer is independent. You can turn any of them on or off, and every layer is measured on public benchmarks. See The layers for what was measured on each.

#Two ways to use it

  • Hosted. We run the gateway. You change your agent's base URL and use a Rakuvo key. See Hosted gateway.
  • Self-hosted. Run the rakuvo gateway yourself, from the installer or a container, and your prompts never leave your network. See Installation and Terminal (CLI).
  • API. Manage organisations, keys, gateway configuration and traces over HTTP, and let your gateways pull their configuration from it. See API reference.

An MCP server is planned and is not part of the first release. See MCP.

#What to expect

Rakuvo is most useful where models fail in the specific ways the layers target. It does not make every model better: on models that already call tools well, the Prompt layer does nothing measurable, and the docs say so wherever that applies.