A verdict per option
Fits, fails or cannot tell, for every requirement, with the value and where it came from.
For AI agents
A shopping assistant that never invents a spec.
Your assistant knows what the shopper asked for. Pulse tells it which products the listings prove fit, which fail, and which say nothing, so it recommends from evidence and says plainly what it cannot tell.
Where you are
Assistants are built to answer.
Ask a language model which laptop has 16GB of RAM and it will pick one, even when the listing never stated the memory. The shopper buys on that answer, and the return reason says "not as described".
"The listing does not say" is a better answer than a confident guess.
What happens
How does it work?
One call, from MCP or HTTP. Nothing to train, nothing stored.
- Ask01
State what is needed
As requirements such as "voltage = 24V", or in plain words; a budget in the words becomes a price requirement.
- Send02
Pass the options
Up to 50 at a time, with whatever fields and text you have. A stated field always wins over text read off a page.
- Answer03
Answer from the verdicts
Each option comes back fits, fails or cannot tell, with the attributes that are missing. The agent says what the evidence shows, and only that.
What you get
What the agent gets
"The evidence does not say"
A first-class answer instead of a guess, with the missing attributes named.
A to-do for the owner
Every unknown is exactly what the option's owner should publish next.
MCP or HTTP
The open-source MCP server for agents, or one HTTP call from your own code.
A worked example
"Laptop under $1,500 with 16GB RAM": one fits, one fails on price, one cannot tell because its listing never states the RAM.
What you get
Before and after
Without Pulse
- Recommends from what it thinks it read
- Fills unstated specs with likely values
- Drops products it is unsure about, silently
With Pulse
- Recommends only what the listing proves
- Says which spec the listing leaves out
- Keeps every product, with its verdict
Before you commit
What it will and will not do
What this does
- Give each option a verdict per requirement, grounded in its own evidence
- Name what is missing instead of guessing it
- Work in any domain whose options carry structured attributes
- Answer the same input the same way every time
What it does not do
- State an attribute the evidence never stated
- Read facts out of free text outside laptops yet; send the fields instead
- Make decisions about people, such as hiring or lending
- Promise a business result. Nothing has been measured on your data
Questions
The ones that decide it.
- Does it rank products for the shopper?
- Not this call: it decides eligibility, requirement by requirement. Ranking a storefront for the live session is the Pulse SDK and Decision API.
- Which agents can use it?
- Claude Code, Claude Desktop, Cursor and Codex through the open-source MCP server, which runs on your machine, and any agent or program that can make an HTTP call. ChatGPT's web app needs a remote MCP server, which is planned.
- Is it a model?
- No. It applies rules to the evidence each option states, so the same input always gives the same verdict, and every verdict shows its source.
Add it to your agent.
Install the MCP server with one line, or call the API from your code. The open-source repository has runnable examples.
Opensource
An API key comes from the Pulse dashboard.