Stimulir

• Infer — Run a chat completion on the Stimulir gateway and return the text. Pass a plain `prompt` (optionally with a `system` instruction) for a one-shot call, or pass `messages` — a full OpenAI-shaped conversation history of {"role", "content"} dicts — for multi-turn use; `messages` wins when both are given. The response streams server-side and progress notifications carry the live text as it generates.
• Speak — Generate spoken audio from text (text-to-speech). Synthesizes `text` through Stimulir's authoritative TTS lane and returns the audio as a playable block plus a short text receipt. Use this whenever the user wants something read aloud, narrated, or turned into an audio message. `provider` defaults to openai; `voice`, `character`, and `model` override the provider's defaults when set. `character` is a voice-style/persona directive (e.g. "Speak with a warm British RP accent
• Speak To File — Generate spoken audio from text and return a hosted URL to the audio file (for uploading/attaching elsewhere, e.g. Slack). Synthesizes `text` through Stimulir's authoritative TTS lane, uploads it to the caller's tenant storage, and returns a signed https URL plus the storage key and a short transcript receipt — as text, not inline audio, so a downstream tool (a Slack/email connector) receives a usable link string. `provider` defaults to openai; `voice`, `character`, and `mod
• Transcribe — Transcribe base64-encoded audio to text through the STT lane. Pass raw base64 or an audio data URL. Input is capped at 25 MB and is sent directly to the authoritative `nova-3` transcription route; it is never written to the MCP server's filesystem.
• Speech To Speech — Send one spoken request and return one spoken response (speech-to-speech). Input must be a 16 kHz mono PCM16 WAV encoded as base64 or a data URL. The returned audio is a playable WAV block from the authoritative realtime model admitted for the caller's workspace.
• Usage — Audit recent Stimulir usage and spend. Fetches the latest usage events (inference calls metered by the gateway) and returns a compact JSON summary: per-model call counts, total event count, and the raw events trimmed to model/provider/mode/timestamp.
• Models — List the model ids available through the Stimulir gateway. Returns the sorted catalog as a JSON array. Use this to discover what `infer` can route to before picking a model.
• List Prompts — List the versioned prompts registered in the Stimulir workspace. Returns a JSON array of prompt metadata (key, name, version, label, type, tags) for every active prompt. Use `get_prompt` with a key from this list to fetch the actual prompt content.
• Get Prompt — Fetch a versioned prompt from the Stimulir registry by key. Returns the prompt's content and metadata as JSON. Pass `label` (e.g. "prod", "staging") to resolve the version carrying that deployment label; omit it for the latest active version. Prompt content may contain {{variable}} placeholders the caller is expected to substitute.

Schemas and invoke metadata: Directory API