Changelog
A week to an entry, newest first: what you can now do, what was fixed, and anything that works differently.
Sarvam Bulbul is a voice engine of its own in the studio, with voices recorded in Hindi, English and seven more Indian languages, in styles from customer care to news reading; some speak English and Hindi together. Sarvam Saaras can transcribe instead of GPT Transcribe, for the Indian languages it takes.
Read more in the docsSend language when you start a call and that one call is held in Tamil, Telugu, Kannada or any other listed language, without changing the agent. Replies are written in the language’s own script, and an agent with a Cartesia voice can move into another language when the caller asks for it.
Upload price lists, policies and FAQs as PDF, Word or text and switch them on for an agent. It answers callers from them, and says so when it does not know rather than guessing.
Read more in the docsOne line of HTML puts a talk button on your site, with no server of your own. It carries a publishable key that can only start calls with that agent, and you set which websites may use it, how long a call may run and how much it may spend a day.
Read more in the docsInvite people to a workspace as an admin or a member. Members build agents and run calls; billing, API keys and webhooks stay with the owner and admins.
Every agent costs the same per minute, whichever voice, model and transcription it uses, and the studio’s voice library lists each engine’s voices by language with a sample to hear first.
Read more in the docsA workspace key (ck_…) lets your server build agents, start calls and read call logs through the API, and is created and revoked in the studio. The docs are at cracklehq.com/docs, with guides, an API reference and worked cookbooks.
An agent’s tools are its own, so one agent’s Tools page no longer lists every other agent’s. Copying an agent copies its tools.
Read more in the docsA short “हाँ”, “okay” or “hmm” while the agent is talking is heard as a listener nodding along, so the caller hears the whole reply.
An agent on Cartesia Sonic can speak any of 44 languages, each reply written in its own script, and every agent has a volume setting beside its speed.
Read more in the docsEach agent’s instructions end with writing rules for its voice: Hindi in देवनागरी, numbers as spoken words, short sentences. On voices that can perform them, a written laugh or sigh is acted out rather than read aloud.
Read more in the docsA call on a workspace with a very large balance could end the moment the caller joined. It now runs to its limit.
Updating a webhook no longer pretends to change its secret: that request is refused. To change a secret, create a new webhook and delete the old one.
Read more in the docsWhen a caller speaks while the agent is talking, it stops and listens, then answers what was just said. A cough, an echo of its own voice or a stray sound no longer counts as an interruption, and the agent knows where it was cut off.
Read more in the docsBuild one in the studio and talk to it in your browser. Sign-up credit is there for your first test calls.