Guides
FAQ
Short answers to common questions.
#Do I need to change my agent's code?
No. Change the base URL, and optionally add the X-Rakuvo-Agent header.
#Which providers and endpoints does it support?
The gateway speaks the OpenAI-style API: POST /v1/chat/completions and POST /v1/responses. Point -upstream at any provider that offers a compatible endpoint. There are compatibility switches for providers that differ (-max-tokens-field, -omit-temperature).
#Does it add latency?
Layers that only inspect or edit a request add very little. Layers that call a model themselves, such as the judge, a summariser or memory extraction, add that call. The trace records the tokens each layer spends so the cost is visible.
#Does it see my prompts?
The gateway does, because it forwards them. What it reports to the control plane is metadata only. See Security and privacy.
#Can I use only one layer?
Yes. Layers are independent. Many teams start with the loop guard in warn-only mode, or the Prompt layer in automatic mode.
#Which layers have been shown to work?
See The layers for what was measured on each. Gains depend on the model, so test on your own tasks.
#Can the gateway run without the control plane?
Yes. The terminal gateway runs on its own with flags and a local trace file. The control plane adds the dashboard, shared configuration, keys, alerts and usage.
#Where is my data stored?
The control plane stores your organisation's metadata in its database (SQLite on a single node). Memory facts are stored in the gateway's memory directory, and copied to the control plane only if you turn on -report-memory.
#How is pricing set?
By plan, per organisation. See Usage and billing, or talk to us.
#Is there an MCP server?
Planned, not built. See MCP.