One gateway between your agent and its model.
Every request passes through the layers on the way to the model, and the answer comes back through the same ones. Touch a layer to see what it does, and switch it off to see where it sits. A control plane holds your configuration, so changing a layer is a setting, not a deploy.
Prompt
One short instruction so models use their tools instead of guessing.
- 1Checks whether the request offers tools.
- 2Adds the tool-use instruction.
- 3In automatic mode, only for models where it helps.
on multi-turn tool calling with gpt-5-mini, and 44% less cost for each correct answer.
Control plane
Layers are settings
Switch layers on and off per organisation. Gateways pull the change within about 30 seconds.
Every call recorded
Every attempt, and the note each layer wrote, with the tokens charged to the layer.
Your own traffic
Cost, latency, failures and layer actions for each agent, as they happen.
Measured on public benchmarks, same questions with and without Rakuvo
more multi-turn tool tasks done right
BFCL · gpt-5-mini · 19.2% to 35.5%more questions answered from long-term memory
LoCoMo · held-out · vs embedding searchfewer wasted tool calls on stuck agents
Loop guard · held-out · no false stopslower cost for each correct answer
BFCL · gpt-5-mini · multi-turnSix layers. Each one fixes one failure.
Switch on the ones you need. Each runs on its own, in the same order every request.
Memory
A temporal graph of what each user has said, with dates and sources, so a returning user is not a stranger.
Context
Keeps long jobs from running out of room. Old results the agent no longer needs are shrunk to a short note, and the agent can fetch them again if it needs them.
Prompt
One short instruction that makes models use their tools instead of guessing.
Model policy
Chooses which model answers each call, cheap first.
Tools
Checks and repairs every tool call, and a small judge can veto one that makes no sense.
Loop guard
Notices an agent going in circles, nudges it, then stops it with an honest report.
Use it from the terminal or the API.
Three steps from nothing to a gateway in front of your agent. Run it yourself and manage it over HTTP.
curl -fsSL <download-address>/install.sh | shexport OPENAI_API_KEY=sk-...rakuvo serve -listen 127.0.0.1:8081 -tool-policy-auto -loop-guardlistening addr=127.0.0.1:8081 guard=trueHow the reliability is evolving.
How each layer works and what we measured on it.
When an AI agent deleted a company's production database
In July 2025 a coding agent on Replit ran destructive commands during a code freeze, then misreported what it had done. What was reported, what it shows about agents, and where a gateway can and cannot help.
Six layers between your agent and the model
What a reliability layer for AI agents actually is, what each of the six does to a request, and why we built it as a gateway you point your agent at instead of a library you import.
Make your agent reliable, and prove it.
We are working with a small number of teams. Talk to the founder about your agent, or read how it works first.