What’s new in Zumu.

Every agent now runs on either of two voice engines: the chained pipeline you tune stage by stage, or a realtime speech-to-speech engine that hears and answers in one model. How the platform works

Connect a Gmail or Microsoft 365 mailbox, then choose how the agent works it: read-only, acknowledge, draft for approval, or send on its own. See the email channel

Import a list, see the cost estimate before launch, and pause or resume at any time. Calling hours follow each recipient’s own time zone, and do-not-call is checked on every dial. See campaigns

When a visitor is signed in to your site, the agent knows who they are, and their account carries into the conversation. See the web channel

Judge each conversation by a model, by a rule or by a number, in packs you switch on per agent. Every verdict cites the turn it came from, and repeated failures cluster so one fix covers the pattern. See quality checks

Add your website or a Google Drive folder beside your documents. Knowledge re-syncs on the schedule you set, and every answer carries its source. See knowledge

Zumu’s own MCP server puts the platform inside the AI tools your team already uses. Connected apps sign in with OAuth, and a disconnect takes effect on the next request. See the MCP server

Each change to an agent’s instructions is kept with who made it, when and from where, drafts included. Any revision can be restored. See the agent studio

When a chat needs your team, an operator takes the thread in one click. Operators set their hours and see their performance. See live operations

On your website the agent can read, point, click, type, choose and go to. It asks the visitor before anything that changes the page. See the web channel

Sign in with Google or Microsoft, or with a password and a 6-digit emailed code. Every session a member holds is listed: end one, or all the others. See security

Connect a Microsoft 365 mailbox and the agent answers email with the same knowledge, tools and memory it uses on every other channel. See the email channel

A dedicated end-of-turn model listens to the sound itself, not a silence timer, so the agent answers when the caller is done and waits when they’re not. See latency

Script the conversations that matter, mock the tools they call, and run them before a change reaches a customer. See the agent studio

Add keys for OpenAI, Anthropic, Google, Deepgram, ElevenLabs, Cartesia or Fish Audio, per organization or per project. Each key is checked with its provider before it’s saved, and the matching provider charge is waived. See your keys

Calendars, CRM, helpdesk, commerce, payments, messaging and data, connected once and shared by every agent. Add any MCP server with standard OAuth. See the integrations

Returning customers are recognized on any channel, with the fields, tags and summary your team keeps. New facts can wait for review before they’re kept. See memory

One script tag puts the agent on any site, and your styles and ours never collide. Visitors type, or switch to voice without leaving the page. See the web channel

The transcript, the recordings, every tool call with what it sent and what came back, the outcome and the itemized cost, together in one record. See conversation records

Beside Google Calendar and Microsoft 365 Calendar, the agent now books straight into Cal.com and Calendly. See the calendars it books into

The same agent and the same memory now answer in messaging apps, beside phone, web, email and SMS. See every channel

When a conversation ends, transfers or its summary is ready, post it to Slack, update the CRM or call your endpoint. See webhooks and automations

Run on a prepaid balance with auto-reload, or on invoice. Every conversation carries its itemized cost. See pricing

See every live conversation and listen to its audio with the transcript beside it. Join late and read the turns you missed. See live operations

Partners put their name, colors, light and dark logos and their own address on the dashboard, and their customers sign in from the partner’s product with partner SSO. See the partner program

Branch an agent to try a change, send part of your traffic to it, and keep what works. See the agent studio

Choose the voice for each agent from ElevenLabs, Cartesia or Fish Audio, then set pronunciation for names, brands and codes. See voices

Noise cancellation and speaker isolation run on every call, so background voices and noise don’t stop the agent mid-sentence. See voice agents

Agents speak 10+ languages and pick up the caller’s language as they speak it. See voice agents

A call moves to a specialist agent inside the same conversation, and what the caller already said moves with it. See voice agents

The caller holds while the agent calls your person and speaks the brief. It connects only when they accept, and a decline returns the caller to the agent with a next step. See transfers

Every agent thinks through a ranked line of models. A request that times out or errors moves to the next one inside the same turn. See reliability