2 minute readImprovedAgent-ready
Price tools now read arguments in the shapes agents send them
Last Price's MCP servers now read a price call sent as text numbers, camelCase names, nulls or one JSON string, and say back what they read. On 1,905 sold cars across 48 agent tool shapes, every car now gets a price on the first call.
Different agent tools send the same price call in different shapes, and Last Price used to refuse the ones it did not expect. Both MCP servers, the hosted one and the npm one (@last-price/tools-mcp), now read those shapes and tell you what they read. Ambiguous values are still refused.
- Numbers sent as text.
"15100","15,100","$15,100"and"USD 15,100"are read as 15100 when the symbol or code matches the call's currency. - camelCase argument names.
currentPriceis read ascurrent_price, unlesscurrent_priceis also sent. - Null on an optional argument. It is read as if the argument was left out.
- Yes/no arguments sent as text.
"true"and"false"are read as true and false. - Arguments sent as one JSON string. They are read as the object. This one applies to every tool.
- Which tools. Apart from the JSON string reading, these apply to the tools that ask for a price and change nothing, such as
compute_price,verify_priceandfind_market_prices. - Each reading is said back. The answer carries one more line, for example:
Read your arguments as: current_price "$15,100" as 15100. Send them this way next time to skip the reading. - Nothing is guessed. A price with two meanings, like
"12.500"(twelve and a half, or twelve thousand five hundred), a symbol for another currency, like"€15,100"on a USD call, and shorthand like"15k"are still refused with the usual error naming the field. - Tools that save settings stay strict. Tools like price guardrails still refuse
"10"for a number and null for a field. There, null can mean "clear it", and a guardrail written wrong is worth a refusal, not a guess. - Measured result. On 1,905 real sold cars, across 48 simulated agent tool shapes, the average shape used to get a price for 66% of cars (30% for the worst shape) and took 2.2 calls per car. Now every one of those shapes gets a price for 100% of cars on the first call.