Read
Tier 1 tools look at markets, your positions, your policy and token contracts. Nothing at this stage can spend, sign or send.
Walls first, cleverness second
A trading agent is only as safe as the limits it cannot talk its way around. Troya is built so the model may propose almost anything, and execute almost nothing on its own say-so.
01 / The guarded path
Read, prepare, then act
There is exactly one road from a question to a filled order, and it has two checkpoints. The app, the MCP server and the REST API all travel the same road.
Prepare & confirm
In the app the assistant drafts a proposal card. Over MCP, Tier 2 tools simulate the order and ask for a confirmationId. Either way Troya Gate runs against your limits before anything is put in front of you.
Execute & audit
You sign in your own wallet, or the Agent Wallet signs inside its caps. Over MCP, execute_intent takes nothing but the confirmationId and checks the policy one last time. The outcome and the reasoning go into the journal.
02 / Shape
Wide eyes, narrow hands
Troya reads broadly: every perp, every Elysium token it can find, pools, funding, contract powers. It acts narrowly: one guarded path, one policy, one journal.
Limits outside the model
Caps live in policy.json on the server, not in a prompt. They are loaded fresh on every call, so no conversation, however persuasive, can edit them.
One-time confirmations
A proposal in the app expires after 10 minutes. An MCP confirmationId lives for 120 seconds and is bound to a hash of its parameters, so changing the size or the market voids it.
Mainnet, with brakes
Real markets and real money. Pause stops your Agent Wallet at once, and an operator kill switch halts every execution path together.
Fenced text
Tweets, token names and descriptions are written by strangers. Troya wraps them as untrusted text, so an instruction hidden in a token name is read as data and never obeyed.
03 / Why it is built this way
The model is a sharp junior. The walls are the desk rules.
Models are persuadable
A language model will happily follow a clever instruction, including one planted in a token description. That is fine for reading. It is not fine for spending, which is why spending never depends on the model agreeing to behave.
Every action has a reason
When the agent proposes something it has to say why, and that reason is stored next to the action. Weeks later you can still see what it was thinking and which rule let it through.
You stay accountable
Troya never pretends the decision was its own. You set the walls, you approve or pre-approve the moves, and the journal shows exactly what was done in your name.