DEVELOPER DOCUMENTATION

One key.
Build your own flow.

Send your Skillies key in the Authorization header. All prices are in INR and share the same prepaid balance.

Authentication

Authorization: Bearer YOUR_SKILLIES_API_KEY
Idempotency-Key: your-unique-request-id

Keep API keys on your server. Reusing an idempotency key will not charge again; a duplicate returns 409 with the original request ID. Audio and transcripts are not retained for replay.

Endpoints

Method / path	Input	Output
POST /v1/chat/completions	messages, max_tokens, stream	Text or SSE chunks
POST /v1/audio/transcriptions	Multipart file: PCM16 mono WAV, 0–30 seconds	Transcript JSON
POST /v1/audio/speech	JSON input, voice: default	24 kHz WAV
POST /v1/converse	Multipart file, optional messages JSON and output_audio=false	Transcript, reply, optional base64 WAV
GET /v1/usage	API key	Balance, latest requests, charges and latency
GET /v1/models	API key	Public Skillies model aliases

Start with chat

curl https://api.skillies.ai/v1/chat/completions \
  -H "Authorization: Bearer $SKILLIES_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"skillies-chat-1","messages":[{"role":"user","content":"Hello"}],"max_tokens":128}'

Give text a voice

curl https://api.skillies.ai/v1/audio/speech \
  -H "Authorization: Bearer $SKILLIES_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"skillies-voice-1","input":"Welcome to Skillies.","voice":"default"}' --output speech.wav

Use skillies-stt-1 for transcription, skillies-chat-1 for intelligence, and skillies-voice-1 for speech. You can combine endpoints in any order.

Billing and errors

Each request reserves a maximum charge, then settles measured usage. Cached input receives the cached rate only when reported. Failed requests are refunded. Response headers include X-Request-Id, X-Charged-Paise, X-Balance-Paise, and X-Latency-Ms. Streaming chat returns billing in its final event. total_ms measures server processing through billing settlement; provider_total_ms isolates model work. auth_ms, reservation_ms and settlement_ms show gateway overhead. response_ready_ms also includes telemetry persistence. Client network and download time must be measured by your application.

401: invalid key · 402: insufficient credits · 409: duplicate/conflicting request · 413: input too large · 502/503/504: service unavailable or timed out.

Real-time sessions

Connect to wss://api.skillies.ai/v1/realtime. Authenticate using the Authorization header, or send a session.authenticate event with your key as the first message. Send mono PCM16 binary chunks at 16 kHz, then a turn.commit event with a unique id. The server returns text.delta, transcript, audio.delta (base64 mono PCM16 chunks at 24 kHz), and turn.done with usage and billing. Send turn.cancel on interruption. Sessions close after four minutes; reconnect and supply messages in session.configure to continue. Speech starts streaming after the text reply is complete. Each submitted turn is limited to 30 seconds.