A pre-trade check for your AI trading agent
A pre-trade check asks one question before an agent places an order: which of the owner’s own written rules would this change put outside range? Allocait answers it from the owner’s rules, holdings and the prices the agent supplies, and leaves the decision to the owner.
Why an agent needs a check outside the model
An AI agent that can call a broker’s order tools usually knows nothing about its owner’s limits. A limit written into a prompt is text the model can misread or drop, and it may not do the arithmetic the same way twice.
A check outside the model reads the rules the owner wrote and does the same arithmetic on every call. The agent asks, reads the answer, and acts on it or leaves it.
What the check states
For a proposed change, Allocait reports facts about the owner’s own rules:
- Which asset-class bands and position size ranges the change would put outside range.
- Which it would leave further outside than they already are.
- Whether the whole portfolio sits inside every rule afterward.
- Which holdings the check could not cover, so a gap is named rather than skipped.
How an agent asks
The same steps work over MCP or REST, signed in through the connector or with a key.
- List the portfolios and read the snapshot, which names the holdings that need a price.
- Get a price for each from a source of your own. Allocait supplies no prices.
- Call check_trades with the proposed trades, each with a symbol, a side and a share count, and a prices map that covers every held and traded symbol.
- Read newly_or_further_outside_count, within_rules and not_covered in the result.
- Decide, and place any order with your own execution tool.
What the check cannot do
It is a check the agent asks for. Allocait places no order, connects to no brokerage and cannot stop an order, so an agent skips it unless its instructions say to call it first. Put the call in the agent’s standing instructions and in whatever workflow places orders.
A result inside every range does not say a change is advisable. It says only where the change leaves each of your rules.
Rules in a prompt, or rules in a check
- In a prompt: simple to write, and read by the model each time. The model decides how to apply them.
- In a check: written once as bands and size ranges, read by a program, and answered the same way on every call. The owner keeps the numbers in a file that people and agents can read.
Where Allocait fits
Allocait is not an execution layer. It sits beside whichever tool places the order and answers one question before it is placed. The rules are the owner’s: Allocait types no band, no size range and no number for a person.
- Setup guide for a coding agent: the steps a coding agent follows to add the server.
- Write your rules as a file: a template with every number blank, and the same rules as a POLICY.md.
- How rebalancing bands work: a minimum and maximum for each asset class.
- Size one position from a stop: the free calculator runs in your browser.
Questions
Write your rules once. Your agent checks them.
Allocait checks a proposed change against the ranges you set, through your AI agent or in the app.
Join the waitlist