Reference

Configuration reference

Every flag of rakuvo serve, grouped by what it controls.

Run rakuvo serve -h for the authoritative list. The same settings can be pushed from the control plane (see API reference) and are validated there.

#Gateway

FlagMeaning
-listenAddress to listen on. A non-loopback address requires -token-env.
-token-envName of the environment variable holding the bearer token clients must send.
-traceFile to write one JSON line per model call.
-upstreamBase URL of the model provider (default OpenAI).
-upstream-key-envEnvironment variable holding the provider key.
-max-tokens-field, -omit-temperatureProvider compatibility switches.

#Prompt and tools

FlagMeaning
-tool-policyAlways add the tool-use instruction.
-tool-policy-autoAdd it only for models where it was measured to help.
-guard, -guard-pass-throughTool-call guard on or off; let a failing call through instead of refusing it.
-verify-modelJudge model. Empty turns Verify off.
-verify-modeveto, gate or both.
-verify-scopesingle (default) or all.
-judge-prompt-fileReplace the judge's system prompt.

#Context

FlagMeaning
-context-keep-observationsKeep this many of the latest tool results in full; mask older ones.
-context-mask-stepMask in jumps of this many results so the cached prompt prefix stays stable.
-context-mask-within-turnAlso mask inside the latest user turn (long single-request agent runs).
-context-max-tokensToken budget for the whole request; drops the oldest whole turns when over.
-context-summarize-model, -context-facts-modelModels for summaries and the key-facts note.
-context-keep-relevant, -context-prune-toolsKeep results relevant to the latest message; prune the tool list.

#Loop, model policy, effort, memory

FlagMeaning
-loop-guard, -loop-warn-onlyTurn the loop guard on; only nudge, never stop.
-loop-max-tool-calls, -loop-allow-repeatHard cap on tool calls; tool names allowed to repeat.
-model-cheap, -model-strongThe two tiers of the cascade.
-model-escalate-on-no-callAlso escalate when the cheap model answered in text.
-reasoning-effort, -reasoning-effort-escalated, -reasoning-forceReasoning effort for reasoning models, and a higher level for retries.
-memory-dirDirectory for per-user memory. Required to enable Memory.
-memory-model, -memory-embed-model, -memory-select-modelModels for extraction, embeddings and evidence selection.
-memory-top-k, -memory-scopeHow many facts to recall; how users are separated.

#Control plane

FlagMeaning
-config-from-apiPull layer settings from the control plane instead of flags.
-api-urlControl-plane URL.
-api-key-envEnvironment variable with your Rakuvo API key (default RAKUVO_API_KEY).
-config-pollHow often to check for new settings.

Layer flags conflict with -config-from-api: settings come from one place.