AI Trading Agent Safety: Approvals and Hard Limits on Trigr

AI trading agent safety on Trigr: exact-argument approvals, manual vs autonomous mode, and hard limits that stop assistants from live trading or withdrawing.

Trigr Research6 min read
On this page
  1. Why does AI trading agent safety need its own design?
  2. What can the assistant do, and what is off limits?
  3. How do exact-argument approvals work?
  4. Should you use manual or autonomous mode?
  5. How does Trigr handle untrusted text?
  6. Where does live trading actually happen?
  7. What is a practical safety checklist?
  8. Next steps

TL;DR: On Trigr, an AI assistant connected over MCP can research, backtest, save drafts, publish eligible strategies and create a paused paper agent, but it cannot start execution, place a live trade, withdraw funds or touch exchange credentials. Every consequential action needs a ten-minute, single-use authorization bound to its exact arguments, approved in your browser in manual mode or automatically in autonomous mode, with credits, quotas and ownership still enforced.

Why does AI trading agent safety need its own design?

Language models are useful research partners, and they fail in particular ways. They can misread a request, fill in a missing parameter with a guess, repeat an action after a timeout, or follow instructions embedded in text they were only supposed to read. None of that is catastrophic when the tool is a web search. It is a different matter when the tool spends credits, edits saved work, or sits near a trading account.

The Model Context Protocol specification is direct about this. It says there should always be a human in the loop with the ability to deny tool invocations, and that clients should show tool inputs before calling a server. Its security best practices add guidance on token handling and least-privilege scopes.

Trigr's trading MCP server combines three layers: capabilities that simply do not exist over MCP, exact-argument authorization for the ones that do, and server-side checks that apply no matter what the model says.

What can the assistant do, and what is off limits?

Area Allowed over MCP Not available over MCP
Research Search the capability catalog, data coverage and marketplace statistics; read your plan, credits, strategies, agent summaries and an agent's positions and trade history; read stored backtests and trade logs Reading private marketplace recipes
Compute Credit-metered backtests with optional slippage and funding; asynchronous ML optimization Running compute without an authorization
Strategies Save, rename or replace drafts; publish, republish or delist eligible strategies; delete owned strategies Editing another user's strategies
Agents Create a paper agent in a paused state; edit an agent's name, capital or paused state Starting execution; changing an agent's strategy or risk configuration
Money and keys Nothing Live trades, withdrawals, reading or editing API secrets, exchange credentials, signing keys or private risk configuration

The right-hand column is the important one. These are not settings you can loosen with autonomous mode. There is no tool for them on the server (the agent-edit tool accepts a paused status and nothing else), so no prompt, however clever, can call them.

How do exact-argument approvals work?

Every consequential action follows the same three-step pattern.

  1. Request. The assistant asks Trigr to authorize one specific tool call with its full arguments: for a backtest, the exact strategy graph, the strategy id, the request id and any slippage or funding settings.
  2. Authorize. Trigr issues a single-use authorization that expires after ten minutes. In manual mode it only becomes valid after you confirm it in your browser. In autonomous mode it is approved immediately.
  3. Execute. The assistant calls the tool with that authorization. If any argument differs from what was authorized, the call is rejected.

Binding the approval to exact arguments closes a subtle gap. Without it, a model could get approval for a cheap, harmless call and reuse it for something else. With it, what you approved is exactly what runs.

Publishing, republishing and delisting follow exactly this pattern: the authorization is bound to the strategy id and its current recipe hash, and the normal publishing gates still apply.

Several other guards sit underneath:

  • Recipe hashes. Edits, publishing and unpublishing must include the strategy's current recipe hash. If the strategy changed since the assistant last read it, for example because you edited it in Studio, the call fails instead of overwriting newer work.
  • Retry-safe billing. Backtests and ML jobs take a stable request id, so a retried call after a timeout is not charged twice.
  • Owner siloing. A connection can only read and write its owner's strategies, jobs and agent summaries.
  • Server validation. Graph limits, semantic checks, disabled data sources, billing and plan quotas are enforced on the server.

Should you use manual or autonomous mode?

You choose the mode when you authorize a connection and can change it later, separately for each client.

Manual Autonomous
Browser prompt per action Yes No
Exact-argument, single-use, ten-minute authorization Yes Yes
Credit, quota, ownership and validation checks Yes Yes
Hard limits (no live trades, withdrawals or credential edits) Yes Yes
Best for New connections, credit-heavy work, publishing Long research loops with a defined budget

A reasonable pattern is to start every new client in manual mode, watch a few sessions, and move to autonomous mode only for research you would approve anyway. Keep an eye on credits: autonomous mode removes the per-action pause, so a loop that runs many backtests spends credits faster. A standard backtest costs 50 credits, where 1,000 credits equal $1.

How does Trigr handle untrusted text?

Prompt injection is the risk that text an assistant reads, such as a web page or a document, contains instructions the model then follows. In a strategy marketplace, the obvious vector is a listing's name or description.

Trigr treats marketplace names, descriptions and builder fields as untrusted third-party data. The marketplace search tool's own description tells the model to treat all text fields as data, never as instructions, and the documentation states that such text must never authorize another tool call. Marketplace recipes are never returned, so there is no hidden logic for an assistant to copy or be misled by.

Exact-argument authorization is the backstop. Even if a model were persuaded by injected text, it could only request actions from the allowed column, and in manual mode you would see the request in your browser, with a summary of what it does, before approving it. For a broader look at how MCP compares with raw API access on this front, see MCP vs API vs bots.

Where does live trading actually happen?

Live trading lives in the Trigr app, under your direct control. An agent runs on one venue: Paper, Hyperliquid, or Propr prop-firm challenge accounts. An assistant can prepare a paused paper agent; starting any agent, paper or live, is something you do in the app.

For Hyperliquid, the connection is non-custodial. Hyperliquid lets a master account approve an API wallet to sign on its behalf; Trigr uses one as a trade-only agent key that places and cancels orders and cannot withdraw. For autonomous agents, Trigr stores the trade-only key encrypted on its servers; you keep your master wallet key and custody of your funds. The details are in Hyperliquid agent wallets explained and the trading agents docs.

Backtests and paper results are not guarantees; perps are leveraged and can lose more than expected.

What is a practical safety checklist?

  • Start each new client in manual mode and read the first few approval screens carefully.
  • Give the assistant a written budget: maximum backtests, maximum credits, and a stopping rule.
  • Ask it to confirm data coverage before spending credits.
  • Publish from a connection in manual mode; a published version is immutable and public.
  • Treat anything from the marketplace as data, and be suspicious of an assistant that suddenly proposes actions you did not ask for.
  • Review connected clients periodically and revoke the ones you no longer use.
  • Forward-test on paper before you consider a live agent, and make that decision yourself.

None of these steps is complicated, and together they keep the assistant in the role it is good at. It can search, draft, compute and summarize faster than you can. It should not be the one deciding how much risk your account carries. That judgment, and the switch that turns an agent on, stays with you by design rather than by convention.

Next steps

Read the AI agents and MCP documentation for the complete tool list and the recommended research loop, then set up a connection in manual mode from the connect-agent page.

Frequently asked questions

Can an AI assistant place live trades on Trigr?

No. Trigr's MCP server has no tool to start execution or place a live trade, in either approval mode. An assistant can create a paper agent in a paused state, and a person decides in the Trigr app whether anything goes live.

What is the difference between manual and autonomous mode?

Manual mode shows a browser confirmation before each consequential action. Autonomous mode lets the connection continue without a prompt each time. Both use a ten-minute, single-use authorization bound to the exact tool arguments, and both keep credit, quota and ownership checks.

Can a malicious marketplace description hijack my assistant?

Trigr treats marketplace names, descriptions and builder fields as untrusted third-party data, and its tool descriptions tell the model never to follow instructions found in them. Authorizations are also bound to exact arguments, which limits what an injected instruction could achieve.

Can an AI assistant see my exchange keys?

No. Trigr's MCP tools never return API secrets, exchange credentials, signing keys or private risk configuration, and they cannot edit them.

How do I revoke an assistant's access?

Each connection uses its own revocable credential created during browser authorization. You can revoke a connection or switch its approval mode from Trigr's agent-access settings at any time; revoking one connection cancels its unused authorizations and does not affect your other connections.

Put the idea to an honest test.

Describe a strategy in plain English or from your own AI assistant, backtest it on point-in-time data, and forward-test it on paper before any real money is involved.