VoiceSplit: voice-first group expense splitting, powered by Vocal Bridge

A voice agent that creates a group, logs who paid what, and tells you who owes whom, built on Vocal Bridge's Voice for App surface.

The problem: splitting a bill is a conversation, not a spreadsheet

Group expenses are messy because they happen in motion. One person grabs lunch, another covers the ride, a third buys the round at dinner, and nobody wants to stop the trip to open an app, tap through a group, and type in each line item. By the time everyone is home, the receipts are gone and the math is a guess.

Naveen built VoiceSplit because the existing apps disappointed him. He first wrote it as an MCP tool so he could manage group expenses through something like Claude Desktop, modeling the group as a graph where nodes are people and edges are who owes what amount. The logic was solid. The interface was still the bottleneck. Typing expenses into a tool is only marginally better than typing them into an app.

What he built

VoiceSplit lets you run the entire group-expense flow by talking. You create a group, add the people in it, log each expense out loud as it happens, and ask for the final settlement, and the agent does the bookkeeping and the math, including who owes whom.

The point is that the whole interaction stays a conversation. You never leave the moment to maintain a ledger. Voice fits because splitting expenses is a talking problem, the kind of thing that happens right at the table, not a form-filling one.

How VoiceSplit uses Vocal Bridge

VoiceSplit is built on Vocal Bridge's Voice for App surface. The voice agent and the backend talk over the same real-time channel, so each spoken instruction maps to a tool call that mutates the underlying expense graph. In the backend logs during the demo, you can watch the calls fire as Naveen speaks: CreateGroup for "Vegas Trip," AddPeople for Naveen and Harish, AddExpense for each line item, and GetGroupInfo to compute the balances.

from vocalbridge import tool

@tool
def add_expense(group_id: str, payer: str, amount: float, label: str):
    """Log an expense and re-split it across the group graph."""
    group = store.load(group_id)
    group.add_edge(payer=payer, amount=amount, label=label)
    group.rebalance()          # recompute who owes whom across the graph
    store.save(group)
    return group.settlement()  # e.g. "Harish pays Naveen $15"

Illustrative. The real implementation uses Naveen's graph-based splitting logic from his MCP tool.

What he did not have to build is the voice layer. Vocal Bridge owns the real-time, sub-second voice channel, the streaming text-to-speech, transcript handling, and the conversational fillers ("getting that set up now") that keep the caller oriented while a tool call runs. He kept the part that was his: the graph model, the settlement math, and the expense store.

What's hard about voice here

A settlement query reads and recomputes the whole group, which takes a beat, and a beat of silence reads as a dropped call. Vocal Bridge's barge-in and filler handling let the agent hold the line and check in naturally instead of leaving dead air while it computes.

That is the honest texture of voice, and handling it gracefully is what keeps the whole flow feeling like a conversation rather than a form.

"After looking at Vocal Bridge, I wanted to turn this idea into a voice agent and, within no time, I was able to do it." Naveen Kumar Vandanapu, builder of VoiceSplit