Prices refresh hourly on weekdays. Each figure carries its own observation date, so a good answer quotes it. A question asked at the weekend gets the newest weekday figure.
The connector is read-only. It looks things up and cannot change anything in your account.
Four questions, four tools
The assistant picks the tool. This is what each question needs.
Each task, the tool it calls and what comes back
Task
Tool
Returns
This week
get_metal_change
Start, end and change
A year
get_gold_silver_ratio
A dated series
Carats
get_gold_carat
Price per gram
A country
get_central_bank_gold
Holdings in tonnes
Before you start
Create a Bullion API account and verify your email address. The connector signs in with that account, so there is no API key to copy.
Check that your Claude or ChatGPT plan, or your workspace administrator, allows custom connectors. Availability differs by plan and changes over time.
Connecting and listing tools is free. Each lookup uses credits from your plan, the same as a REST call. See the pricing page for your allowance.
Never paste an API key into a chat. The connector does not need one, and anything you paste into a conversation can be kept by the app.
Add the connector
Both apps ask for the same server address. Add it as a custom connector (Claude) or a custom MCP server (ChatGPT), name it Bullion API, and choose OAuth sign-in if asked.
https://api.bullionapi.dev/mcp
Menu names move between app releases, so the click-by-click steps are kept in one place: Claude and ChatGPT on the MCP reference page. It also covers sign-in problems.
Approve access
The app opens a Bullion API page. Sign in, check that you started the connection, and approve. The page shows the app's name as unverified, because any app can choose its name. It asks for these read permissions:
Market data: Read precious-metals market prices and related market data.
Macroeconomic data: Read macroeconomic indicators and time series.
Your private portfolio: Read your account’s private portfolio and transaction history. This does not allow changes.
Prices and central-bank gold need only the first two. Turn the connector on for your conversation, then try the questions below.
Four tasks, with the tool call behind each
Each task shows a question you can paste, the tool call the assistant makes, and the shape of what comes back. The values are placeholders because prices change every hour. The wording of the assistant's own reply varies from run to run, so none is reproduced here.
1. Gold in pounds this week
Ask: How has gold moved in pounds over the past week?
change_percent is relative to the start price. A good answer names both dates and the currency and unit, and rounds to two decimals only when it writes the number out.
2. The gold-silver ratio over a year
Ask: Chart the gold-silver ratio over the past year and tell me where it is now.
The assistant works the two dates out from today's date. The series has one point for each UTC day with both prices. Ask for a chart and the assistant plots those points; the values stay decimal strings in the result.
3. Gold by carat, per gram
Ask: What is 18-carat gold worth per gram in pounds today?
The result lists every carat from 24k down to 6k, and as_of says when the underlying price was observed. Each value has eight decimal places, so the assistant rounds for display only.
4. Central-bank gold for a country
Ask: How much gold does the central bank of Germany hold, and is the figure up to date?
Check stale before trusting how recent a figure is, and holdings_basis to see whether the weight was reported or derived. Each series carries the publisher's reuse terms, so ask the assistant to include them when you quote the figure. The block above is abridged: each series also carries a label, a measure and its method. Leave the country out and the tool returns the latest figure for every covered country instead.
Check the answer yourself
Both apps show that a tool ran and let you open it. Menu labels vary, but the tool name and its arguments are what to look for.
The result holds the exact decimal strings. If the assistant's number differs from the tool result beyond rounding, ask it to quote the result.
If a tool returns no_data, the assistant should say so and not estimate. Prices are informational and are not financial advice.
The status is the one the matching REST call returns. Inside a tool result, the assistant sees the error code and message in this form, and should explain it:
{
"status": "error",
"error": {
"code": "no_data",
"message": "No stored data is available."
}
}