Case study · Customer support automation
The support agent spent more context on tool definitions than on the ticket.
A support team's AI agent connected to a dozen MCP servers and drowned every prompt in tool schemas. Moving to onemcp's Code Mode put the context back on the customer's problem — and cut cost and latency per reply.
- Industry
- SaaS customer support
- Team
- ~15-person support org
- Stack
- Zendesk · Stripe · Postgres · internal orders API · Slack
# At a glance
- ~90% lower per-request tool overhead
- ~15,000 → ~500 tokens of tool schema per call
- Lower cost and latency on every reply
Illustrative scenario based on common onemcp deployments. The team is a composite, not a named customer, and the figures are representative rather than measured.
Background
A SaaS support team built an AI agent to triage and answer tickets. To do real work, it needed reach: Zendesk for the ticket, Stripe for the charge, Postgres and an internal orders API for account state, and Slack to escalate. Each of those was an MCP server, wired straight into the agent.
The agent was useful from day one. It was also slow and expensive in a way the team could not fully explain — until they looked at what was actually in each prompt.
The challenge
Every MCP server the agent connected to injected its entire tool schema into the context on every request. With a dozen servers connected, that was hundreds of tool and parameter definitions riding along with each ticket — whether or not the current ticket needed them.
The cost showed up three ways:
- Wasted context. Roughly 15,000 tokens of tool boilerplate led every prompt before the customer's actual question. On long threads, definitions crowded out the conversation itself.
- Higher bills and latency. Those input tokens were paid for and processed on every single turn, multiplied across thousands of tickets a week.
- More room to go wrong. A larger, noisier prompt gave the model more ways to pick the wrong tool or lose the thread of a multi-step ticket.
The approach
The team put all of the agent's servers behind a single onemcp portal and switched the agent to talk to it through Code Mode.
Instead of pre-loading every definition, the portal exposes three meta-tools:
search— find the tools relevant to the ticket in front of the agent.describe— pull the exact signature for just those tools, on demand.execute— run a short script that calls them, in a single round-trip.
So a refund ticket loads the Stripe and orders tools it needs, and nothing else. The full catalog is still available — it simply isn't dragged into every prompt.
Results
- ~90% less tool overhead per request. Per-call tool schema dropped from roughly 15,000 tokens toward ~500. The saving repeats on every turn of every ticket.
- Cheaper and faster replies. Fewer input tokens per call meant a lower bill and a quicker first response, with no tools removed from the agent's reach.
- Steadier multi-step handling. With the context focused on the ticket rather than the toolbox, the agent stayed on track through refunds, account lookups, and escalations.
The agent got faster and cheaper the day we switched, and we didn't remove a single tool. It just stopped carrying all of them into every message.