TL;DR: Trigr's Studio copilot takes a plain-English request such as "BTC trend strategy with a volatility filter", builds a valid node graph from the capability catalog, applies it to your canvas, and backtests it on the real engine. It is fast and useful, but you still need to check five things before trusting the result: the graph itself, the cost assumptions, the trade count, the trade log, and how many experiments it took to get there.
What is an AI trading copilot?
An AI trading copilot is an assistant built into a strategy tool that can read and change your strategy, not just talk about trading. The difference from a general chatbot is grounding. A chatbot without tools will describe an RSI strategy and may even invent a win rate. A copilot connected to a real builder and engine can only use indicators that exist, and its numbers come from an actual backtest.
In Trigr, the copilot sits beside the Studio canvas. Studio strategies are node graphs with a fixed shape: exactly one TRIGGER, any number of FILTERs, a SIGNAL that combines them, a RISK node for sizing and exits, and an EXECUTE node. The pillar guide to the no-code trading strategy builder explains each node. The copilot works inside that structure, which is what keeps its output checkable.
How does the copilot turn English into a graph?
The copilot works as an agent with a small set of tools. In a typical turn it:
- Reads the current graph, including the market, timeframe, nodes and any validation issues.
- Searches the capability catalog, the authority on indicator ids, data-source ids, parameter ranges, assets and timeframes. If it wants a funding-rate filter, it looks up the exact source first.
- Checks data coverage for an asset and data source before spending a backtest on a series that might not cover that market.
- Applies an edit as a complete, ordered node list. The edit lands on your canvas immediately, and the whole turn can be undone.
- Dry-runs validation when it is unsure a combination is admissible.
- Runs a backtest of the current graph and reads back the headline metrics.
- Diagnoses a zero-trade result with a per-condition pass-rate check that shows which condition never fires.
When a request names only a market or an edge, such as "SOL breakout", the copilot does not interview you. It opens with a short list of working assumptions (market, timeframe, how it read the edge, risk defaults), builds a catalog-supported baseline, names it, and measures it. You refine by replying. If a message has no actionable intent at all, such as "make me money", it asks one clarifying question instead of inventing a goal.
What happens in a longer research session?
For open-ended asks such as "find a good ETH strategy", the copilot switches to a research loop and narrates it:
- It states a short plan of the hypothesis families it intends to try.
- It builds and backtests a baseline, then posts the headline numbers with one line of interpretation.
- It iterates in labelled experiments (Experiment 1, 2, 3...), saying what it will change before each run and what the numbers showed after.
- Every few experiments it posts a small scoreboard of the variants so far.
- It finishes by leaving the best-measured configuration on the canvas and naming the winning experiment, with honest caveats about sample size and the number of trials.
Each experiment is stored in the strategy's version history, and any version can be restored. The case for working this way, rather than saving each variant as a new strategy, is covered in iterating on a strategy with AI experiments.
The copilot works to its own default bar: positive net return, a Sharpe ratio of at least 1, maximum drawdown within 20%, and at least 20 trades on the full-history backtest. It is told to present that as its working target, not as your requirement, and to stop honestly when a few experiment families fail to move toward it.
What guardrails does the copilot work under?
A copilot that edits a strategy you might later run with money needs constraints. The important ones in Trigr are:
| Guardrail | What it prevents |
|---|---|
| Catalog-only ids | Invented indicators or data feeds that do not exist |
| Server-side graph validation | Structurally invalid graphs, such as two triggers or no exit |
| Declared changes only | Silent edits: an undeclared removal or risk tweak is rebased back to the previous state |
| No unrequested leverage or size increases | The copilot raising your risk on its own initiative |
| Honest reporting rule | Rounding a loss into a flat result, or claiming an edit landed when it failed |
| No deploy tool | The copilot cannot start an agent or place a trade |
The "declared changes only" rule deserves a note. The copilot must describe every change in its edit summary. If an edit quietly drops a filter or changes a stop without saying so, the server restores the previous state. What you read in the chat is what actually changed.
Some requests are simply out of scope, and the copilot should say so instead of faking them. Pairs and spread trades are not supported in a single-asset graph; it will suggest trading one leg with the other as a cross-asset filter. Custom exit rules such as "exit when RSI is overbought" are not supported; exits are RISK controls (take-profit, stop-loss, ATR trailing stop, a time stop, or exit on signal flip). Two independent entry triggers need two strategies.
What should you check before trusting the result?
A copilot compresses hours of clicking into minutes. It does not remove the need for judgment. Before you act on anything it produces, check these five things.
1. Read the graph, not the summary
Open the canvas and read each node. Confirm the market and timeframe are what you meant, that cross-asset filters point at the right asset, and that the RISK node's direction, size, leverage and exits match your intent. A cross-asset filter that silently reads the primary market instead of BTC is the kind of mistake that looks fine in prose.
2. Add realistic costs
Trigr applies trading fees and the builder fee to every backtest. Slippage and funding are opt-in and off by default, so a first result is labeled gross. Add a flat slippage in basis points per fill and historical or flat funding before comparing variants. Slippage here is a flat assumption, not an order-book model. The guide to slippage and funding in perp backtests shows how much this can move a result.
3. Count the trades
A Sharpe ratio from 12 trades is an anecdote. Look at the trade count and at how the trades are spread over time.
4. Page through the trade log
Check whether the profit comes from a handful of trades or one unusual month. Also check the audit trail, which flags which inputs were real and which were simulated.
5. Count the experiments
If the winner came out of 15 experiments, some of its edge is selection. Studio shades the trailing 25% of the backtest as a recent-period diagnostic, but that is not a holdout: standard backtests use the full available history. For evidence the search did not touch, run ML optimization, which reports out-of-sample Sharpe from anchored, purged walk-forward validation along with the Deflated Sharpe Ratio (Bailey and López de Prado) and PBO, or forward-test the frozen strategy on a paper agent.
Backtests are not guarantees; perps are leveraged and can lose more than expected.
Copilot in Studio or your own assistant over MCP?
Trigr supports both, and they share the same engine and version history.
- Studio copilot: built in, sees the canvas, edits land live, nothing to configure. Good when you want to build visually and watch changes happen.
- Your own assistant over MCP: ChatGPT, Claude, Claude Code or Codex connected through the Trigr MCP server. Good when you already do research in that assistant, want it to combine Trigr with other tools, or want exact-argument approvals on each consequential action. MCP experiments and copilot experiments on the same strategy appear in the same history.
For the underlying mechanics of tool-using models, Anthropic's tool use overview and the Model Context Protocol introduction are good primers.
What does it cost?
Copilot edits and backtests are paid for in credits. A standard backtest costs 50 credits, where 1,000 credits equal $1, and cache hits are free. The zero-trade diagnostic does not spend credits. The Free plan includes 3,000 one-time trial credits; Trader and Pro add monthly credits, and details are on the pricing page.
Next steps
Open Studio, describe one idea in a sentence, and let the copilot build a baseline. Then work through the five checks above before you run a second experiment.