TL;DR: "AI trading platform" covers five very different kinds of tool: plain chatbots, assistants that write strategy code, automated signal bots, copilots built into trading apps, and platforms your own assistant connects to over the Model Context Protocol (MCP). The useful way to compare them is by three questions: does the AI see real data, are its numbers produced by a real engine, and what is it allowed to do to your account. MCP-connected platforms such as Trigr aim to give the assistant real tools while keeping hard limits on execution.
What counts as an AI trading platform?
The label is used for almost anything with a language model or a model of any kind attached. For a trader, that vagueness is a problem, because the tools behave very differently. A chatbot that confidently quotes a Sharpe ratio it never computed and a platform that runs a real backtest can look identical in a chat window.
Three questions cut through most of the marketing:
- Data: does the AI work from real, timestamped market data, or from what it remembers?
- Computation: are the numbers it reports produced by a real engine you can inspect, or generated as text?
- Authority: what can it do to your account, and who approves it?
The five types, side by side
| Type | Real data | Real computation | Account authority | Typical risk |
|---|---|---|---|---|
| Plain chatbot | No, unless given tools | No | None | Invented statistics |
| AI writes your code | Only what you load | Yes, if you run it | Whatever your code has | Bugs you do not notice |
| Automated "AI" signal bot | Yes | Yes, but opaque | Often full trading rights | Black box with your capital |
| Copilot inside a platform | Yes, the platform's | Yes, the platform's engine | Limited to the product | Depends on the platform |
| Assistant connected over MCP | Yes, via tools | Yes, the platform's engine | Only what the server exposes | Depends on the server's limits |
1. Plain chatbots
A general assistant with no tools is a good tutor. It can explain funding rates, critique a hypothesis, or outline how a mean-reversion strategy might work. It cannot see today's prices, and if you ask it for a backtest result it may produce plausible-looking numbers that no engine ever calculated.
Use it for ideas and education. Never treat a number from a tool-less chat as evidence. The post on why most AI trading bots fail covers hallucinated edges in more detail.
2. AI that writes your strategy code
Coding assistants can write a strategy in Python or Pine Script in seconds. This is powerful: the code runs on a real engine and produces real numbers. The risk moves from invented numbers to subtle bugs, especially look-ahead bias, where a misaligned timestamp lets the strategy see the future, and missing costs.
The AI will not warn you about a bug it wrote. You need to read the code, or run it on an engine whose conventions prevent those mistakes. The comparison of Python backtesting libraries and Trigr looks at that trade-off.
3. Automated "AI" signal bots
Many products sell automated trading "powered by AI" with little detail on the model, the data or the backtest. Some are sound; many cannot be evaluated from the outside. They usually need trading rights on your exchange account, sometimes broad API keys.
Before using one, ask what data it uses, how its track record was produced, whether live performance is shown separately from backtests, and exactly what its API permissions allow. The Hyperliquid bot checklist applies to any automated bot.
4. Copilots built into a platform
Some trading platforms include their own AI copilot that drafts or edits strategies inside the product. Because the copilot uses the platform's data and engine, its outputs are real results, not text. The limits are the platform's: what data it covers, how honest its backtests are, and what the copilot can change.
Trigr's Studio has a copilot of this kind. It drafts and edits the strategy node graph from plain English, runs backtests, and iterates in labelled experiments (Experiment 1, 2, 3) inside one strategy's version history, so you can compare variants and restore any version.
5. Assistants connected over MCP
The Model Context Protocol is an open standard that lets AI applications connect to external tools and data. Major assistants support it: OpenAI documents MCP connectors in its API, and Anthropic documents how to add MCP servers to Claude Code.
With MCP, you keep the assistant you already use and it gains real tools from a platform: search real data, build a strategy, run a backtest, read the actual trade log. The platform's server decides what the assistant can and cannot do. This combines the flexibility of a general assistant with the real computation of a platform, provided the server's limits are well designed.
Why do the limits matter most?
Language models are good at producing confident text and occasionally wrong about facts. A trading setup should assume the AI will sometimes be wrong and make that harmless. In practice that means:
- Separating research from execution. An AI that can draft and test is useful; an AI that can also send live orders without review is a single point of failure.
- Approvals bound to exact actions. Approving "run a backtest" should not implicitly approve a different backtest, or a publish, or a change to an agent.
- No withdrawal rights. No research tool needs the ability to move funds.
- Verifiable outputs. Every number should link back to a stored run you can open.
These are also the questions to ask any AI trading platform, whichever category it falls in.
How does Trigr's MCP server handle this?
Trigr runs one remote MCP server at https://trigr.xyz/mcp that works with ChatGPT, Claude, Claude Code, Codex and other MCP clients. Setup uses browser OAuth, which creates a revocable credential, so there is no API key to copy. The connect an AI assistant page walks through it, the MCP pillar article explains the full tool set, and the AI agents and MCP docs list every tool.
What the assistant can do:
- search the capability catalog of assets, timeframes, indicators, data sources and parameters;
- search marketplace metadata and verified statistics;
- save and edit strategy drafts, guarded by a recipe hash so concurrent edits fail safely;
- run credit-metered backtests, with optional slippage (flat bps per side) and funding (historical or a flat per-8h rate), and record variants as labelled experiments on an owned strategy;
- run asynchronous ML optimization, which reports a leak verdict, out-of-sample Sharpe, the Deflated Sharpe Ratio and PBO;
- read stored backtest runs, including runs made in the web app, and their full trade logs for free;
- publish eligible strategies and create a paper agent in a paused state.
What it cannot do: start execution, place a live trade, withdraw funds, or edit secrets and exchange credentials. The server never returns API secrets, exchange credentials or signing keys.
Approvals: each connection runs in manual mode, with a browser confirmation before each consequential action, or autonomous mode, where the connection may continue without prompts. Both use a ten-minute, single-use authorization bound to the exact tool arguments, and you can change the mode per connection. The article on what an AI assistant can and can't do on Trigr covers the details.
The backtests the assistant runs follow the same conventions as the Studio: point-in-time data, next-bar-open fills, fees always applied, and slippage and funding opt-in with a first result labeled gross.
A note on ChatGPT: full MCP write actions currently need a Business, Enterprise or Edu workspace on the web, and other plans may be limited, so check OpenAI's current connector availability. MCP works on every Trigr plan; credits and saved-strategy limits still apply.
When is another type of tool the better fit?
- You want to learn: a plain chatbot is a patient tutor, as long as you do not trust its numbers.
- You want unlimited flexibility: an AI coding assistant plus your own Python engine can express anything, if you check the code.
- You chart across many markets: a charting platform with its own scripting may suit you better; see the TradingView Strategy Tester comparison.
- You want Hyperliquid perps with real data and hard limits: an MCP-connected platform such as Trigr fits that job.
Whatever you use, AI does not remove market risk. Backtests are not guarantees; perps are leveraged and can lose more than expected.
Next steps
If you already use ChatGPT, Claude or Codex, connect it from the connect an AI assistant page and ask it to search the catalog and run one gross and one net backtest of a simple idea, then read the trade log yourself.