Commands
DocsCLIinflexa setup

inflexa setup

Install, authenticate, and start CLIProxyAPI and Postgres (Docker or Podman); optionally configure embeddings

Usage

inflexa setup [options]

Options

Option Description Default
--no-auth Skip the provider authentication step
--no-start Set up only; don't start the proxy or Postgres containers
--no-postgres Skip the Postgres provisioning step
--force Re-pull images even if they are already cached
--no-validate Skip the network probes that check an answered endpoint, model, and credential source. Local GGUF verification still runs
--yes Batch mode: never prompt — every unanswered question takes its default or fails. Implied when there is no terminal
--config <path> YAML answers file carrying any of the batch-mode answers; a flag overrides the file's answer for that question
--connection <mode> How inflexa reaches models: cliproxy|direct (default cliproxy)
--provider <name> With cliproxy: the account kind to sign in — gemini|openai|claude|qwen|iflow (interactive runs only, the sign-in needs a browser). With --connection direct: the vendor slug recorded for the endpoint, e.g. anthropic, openai, deepseek
--base-url <url> Direct connections only: the endpoint inflexa sends model requests to, /v1-terminated
--protocol <wire> Direct connections only: the endpoint's wire protocol — anthropic|openai-compatible (inferred from the provider when omitted)
--model <id> Model id to pin for all model roles; required for a direct connection in a non-interactive run. Under cliproxy it is checked against the account only once the proxy is up, so a rejected id fails setup after the stack is provisioned
--auth-env <var> Direct connections only: name of the environment variable holding the endpoint credential (never the credential itself)
--auth-command <cmd> Direct connections only: command whose output supplies a refreshing endpoint credential
--auth-scheme <scheme> How the credential is sent: x-api-key|bearer. Required with --auth-env or --auth-command
--auth-format <format> How --auth-command's output is read: raw|exec-credential (default raw)
--postgres-user <user> Postgres role the harness connects as
--postgres-password <password> Password for that Postgres role (visible in argv — prefer the --config file)
--postgres-port <port> Host port the Postgres container publishes
--postgres-database <db> Postgres database the harness uses
--postgres-host <host> Host the harness reaches Postgres at
--resource-share <pct> Percentage (1-100) of this machine's CPU and memory sandboxes may use; persisted as the absolute budget
--embeddings <mode> Configure embeddings non-interactively: local (built-in bge-small model)|api-key|off
--embeddings-url <url> api-key embeddings only: the embedding endpoint's base URL
--embeddings-model <id> api-key embeddings only: the embedding model id
--embeddings-gguf <path> Local embeddings only: path to your own GGUF model file instead of the built-in one
--refs <ids> Reference data to download: recommended|all|<comma-separated dataset ids>. The value is the download consent
--sandbox <variant> Sandbox image variant to pull: python|python-r. The value is the multi-GB download consent; nothing is pulled without it
--runtime <runtime> Container runtime to provision on: docker|podman. A hard gate — setup fails rather than falling back when it is not ready