Wakili

Getting started

View raw .md

Wakili documentation

Wakili lets you compose a production voice AI agent from your own STT, LLM, and TTS provider choices, then call it over a REST API and a WebSocket — no telephony infra, no streaming pipeline, no per-provider SDKs to wire up yourself.

You configure an agent (a system prompt + a provider/model choice for each of speech-to-text, the LLM, and text-to-speech) in the Wakili dashboard, generate an API key, and your app talks to two endpoints for a live call:

  • POST /agents/{agent_id}/calls — start a call, get back a one-time WebSocket URL.
  • WS /calls/{call_id}/talk — stream mic audio in, get transcript/reply-text/reply-audio events back.

The same API key and the same balance also cover non-realtime work — one-shot LLM completions, file transcription, and speech synthesis — plus call.ended webhooks, so you can run an entire voice product on Wakili without wiring up a second provider for the text-only steps around a call.

Everything else (billing, usage, recordings, per-call cost breakdowns and metrics) is queryable from the same API.

Where to start

  • Quickstart — create an agent, get an API key, make your first call. ~5 minutes.
  • Authentication — API keys vs. dashboard login tokens, and how to keep keys server-side.
  • The talk WebSocket — the streaming protocol: what you send, what you get back.
  • Agents API — create/list/update/delete agents, and the realtime pipeline knobs (barge-in, endpointing, language lock, model tier).
  • Calls API — start calls, list history, fetch the frozen transcript schema, per-call metrics, cost breakdown, and recordings.
  • Batch LLM API — POST /llm/completions: pick a model, messages in, text or schema-validated JSON out. No agent.
  • Batch STT API — POST /stt/transcribe: pick a provider, upload an audio file, get text + timestamps + optional diarization. No agent.
  • Batch TTS API — POST /tts/speak: pick a provider and voice, text to a streamed audio file, with a server-side cache. No agent.

The batch endpoints bill the underlying model's list price plus a flat 10% routing fee — see each page.

  • Webhooks — register an endpoint, get a signed call.ended POST instead of polling.
  • Next.js integration guide — a complete server-to-server example (push-to-talk in the browser, API key never leaves your server).

Base URL

All REST and WebSocket endpoints are served from the same host, with no path prefix:

https://api.wakili.dev

Locally (via docker compose up from the repo root) the backend listens on http://localhost:8010; running the backend directly with uv run uvicorn app.main:app --reload it's http://localhost:8000.

Reading these docs

Every page here has a Copy as Markdown button — it copies the page's raw Markdown source, ready to paste into an LLM. You can also fetch the raw source directly by appending .md to any doc URL, e.g. https://wakili.dev/docs/quickstart.md — no HTML, no JS, just the Markdown, so an agent or script can curl it directly.