Page three of the quote, under "AI Assistant Module," lists three lines: Tool Use integration, Agent workflow automation, MCP Server build. Total NT$680,000. The owner called and asked whether this was one thing billed three times. It isn't. These are three layers. Each one lets the system do more on your behalf, and each one lets it cause more damage. Get them confused and you either pay Agent money for a chatbot that only looks things up, or you think you bought a lookup feature and it emails 40 customers at 2 a.m.
One restaurant, three terms
An LLM is a waiter with a beautiful vocabulary: he can describe every dish, but he stands there and nothing reaches the kitchen (covered in a separate post). Tool Use hands him an order terminal, so tickets finally get through. Agent lets him decide on his own whether to chase a late table, swap a dish, or stop and fetch the manager. MCP is every kitchen agreeing on the same order-terminal socket, so swapping floor staff doesn't mean rewiring.
Tool Use / Function Calling: the model knocks, you open the door
The model cannot reach your database. Tool Use (function calling at OpenAI, tool use at Anthropic) works like this: you describe your functions and their parameters in JSON Schema, the model decides which one fits and what values to pass, and it returns a structured call request. Your own code does the executing. The model proposes; your Laravel app acts.
A bubble tea chain with 12 stores. A manager asks in LINE how many bags of brown sugar pearls the Da'an branch has left. That becomes checkInventory(store_id, sku), and your code pulls quantity from store_inventories. Permissions and audit live at exactly this layer: does the function verify that store_id is a branch this manager actually runs? Does every lookup land in audit_logs with a user and timestamp? The moment you add createWorkOrder() or sendEmail(), one wrong parameter becomes one wrong work order — so every write tool needs a parameter allowlist plus a confirmation step.
AI Agent: the layer that decides how many laps to run
An Agent wraps Tool Use in a loop: call, read the result, decide the next step, call again, until it judges the job done. The parts are planning, memory, and a stopping condition. The third is where things go wrong — the waiter decides for himself whether to chase the kitchen, which saves you from micromanaging and also means he might chase seven times.
An industrial parts trading company receives a 30-line BOM enquiry. The Agent checks each item against products, pulls missing items from supplier_prices, calculates tax and freight, writes into quotations and quotation_items, and mails a draft to sales. You need four gates: a maximum call count per task (say 60), a token budget ceiling, human review before anything hits quotations, and a rollback on failure.
Now the money. Say one task averages 8 model calls, 48,000 cumulative input tokens and 4,000 output tokens. At Sonnet-tier rates of US$3 / US$15 per million tokens (official pricing) that's roughly US$0.20, about NT$6.5. Two thousand tasks a month lands near NT$13,000. Skip the retry ceiling and average two extra loops and the same volume becomes NT$39,000 a month — and you usually notice in month three.
MCP: a standard socket, not a standard guard
MCP is the protocol Anthropic open-sourced in late 2024 to standardise how an AI client connects to your tools and data sources (official site, specification). Before it, you wrote one integration for Claude, another for a different vendor, another for your internal chatbot. Now you wrap the system as one MCP Server and any MCP-capable client can plug in. The socket is standard; the kitchen hasn't changed.
A medical aesthetics clinic wraps read-only queries over appointments and treatment_records as an MCP Server, with identifying fields masked. Reception uses product A today, marketing switches to product B next quarter, and nobody rewrites the integration. But MCP governs how you connect, not who may connect — hang a "delete patient record" tool off it and it will happily delete.
Side-by-side
| Term | One-line definition | Analogy | Decides on its own? | Who is accountable | How it reads on a quote |
|---|---|---|---|---|---|
| Tool Use | Model proposes which function to call; your code runs it | Punching a ticket into the terminal | Picks the tool, not whether to act | Your backend | "AI query integration" |
| AI Agent | A loop that decides how many calls and when to stop | Waiter chasing and amending orders | Yes, including retry and give-up | Whoever set the stopping condition | "Intelligent agent / workflow orchestration" |
| MCP | Open standard letting any AI client reach your tools the same way | A shared terminal socket | No, it is only a socket | The tools hanging behind it | "MCP Server build" |
Three misconceptions worth killing
- "We wired up an API, so it's an Agent." Without a loop and a stopping condition it is automated lookup. It should be priced as Tool Use.
- "More autonomy is better." Autonomy should be inversely proportional to the cost of being wrong. Inventory lookups can run unattended; issuing discount codes, changing order totals, or emailing customers keep a human gate.
- "MCP means it's secure." MCP is a transport and description standard, not an authorisation system. Scoping, field masking, and audit trails are still yours to build.
What these three words do to the price
- Read-only Tool Use (3–5 tools, permission checks, audit log): NT$120,000–180,000, 3–4 weeks
- Adding write tools and human review gates: +NT$80,000–150,000, 2–3 weeks
- Agent loop (stopping conditions, cost ceilings, failure rollback): +NT$150,000–250,000, 3–5 weeks
- Wrapping existing tools as an MCP Server: NT$60,000–120,000, 1–2 weeks
- Token spend is separate — budget NT$5,000–40,000 a month and demand a usage dashboard
Five questions to ask your vendor
- Is this Tool Use only, or is there an Agent loop? What is the stopping condition and maximum call count?
- Which tools are read-only and which write? Which tables do the write tools touch?
- Is the permission check in the prompt or in my backend? (Only one answer is correct.)
- How do I see token usage? Are there per-task and monthly ceilings, and does it halt when hit?
- If MCP fades in three years, how much of my integration has to be rebuilt?
Decision checklist
- ☐ I can name every table this feature reads and writes
- ☐ Every write tool has confirmation or human review
- ☐ Permission checks sit in the backend, not the prompt
- ☐ Tool calls land in an audit trail with user, time, parameters
- ☐ The Agent has a call-count and token budget ceiling
- ☐ Failed tasks roll back and leave no half-written rows
- ☐ There is a daily token spend dashboard
- ☐ The contract states who pays token costs and how overruns work
- ☐ There is a kill switch that needs no redeploy
FAQ
We only want customer service lookups. Do we need an Agent?
No. For single-step tasks like checking an order or stock level, Tool Use is enough, easier to debug, and cheaper. Paying an extra NT$150,000 for a loop that runs one step is waste.
Is MCP worth adopting now?
If you expect two or more AI clients touching the same data, yes. With one internal system, integrate directly first, keep the tool interface clean, and wrapping it in MCP later usually takes under two weeks.
How bad can runaway Agent spend get?
On the figures above, a normal month is around NT$13,000. With no ceiling and unlimited retries you will eventually hit API rate limits, but burning tens of thousands of NT dollars in a single day is entirely possible.
Want these three layers separated before you sign?
ScriptWalker builds Laravel and Flutter projects for Taiwanese SMEs. Send us the quote and we will mark which lines are Tool Use, which are Agent, and which are billed twice.
- Email: [email protected]
- Phone: 0916-224-047
- LINE: @ufv9089p