MCP connector
What to ask for
The eight guided workflows your assistant can run, with prompts that work — and what it will ask you before it starts.
Updated 2026-08-26·Rev. 2026.06
You do not need special syntax. Ask in a sentence, name the thing you mean, and let the assistant confirm before it writes.
The eight guided workflows
Spera ships eight procedures your assistant can follow. Most clients surface them as selectable prompts; in the ones that do not, simply asking for the task loads the same procedure.
| Workflow | Ask it something like |
|---|---|
| Connect and orient | "What Spera projects do I have?" |
| Build a strategy | "Build an EMA(20)/EMA(50) crossover on BTCUSDT 1h with a 2% stop." |
| Build a module | "Turn my entry logic into a reusable module." |
| Build a custom node | "Make a custom node for a volatility-adjusted position size." |
| Run a backtest | "Backtest that over 2024 and show me the drawdown." |
| Deliver a strategy | "Finish this one — test it, document it, and write up what it does." |
| Update project knowledge | "Record why we switched to the 1h timeframe in the project notes." |
| Document an artifact | "Write the docs for my Trend Rider strategy." |
Each carries the actual method, not just the goal: which revision to pin, what evidence to attach to a claim, and when to stop and ask rather than guess.
What it will check before writing
It will ask which project you mean
If you have several projects and more than one plausibly fits the edit, the assistant is required to stop and ask you to pick — by name and ID — rather than choosing for you. That is deliberate. Creating a new strategy needs no existing project, so a create request never gets blocked on this.
It also pins an exact revision before editing. If the artifact moved since it last looked, it re-reads instead of overwriting your change.
Asking for evidence
The workflows distinguish between what was measured and what was assumed. Useful follow-ups:
- "Did that compile, or are you assuming?"
- "Show me the trades behind that Sharpe."
- "Which revision did you actually test?"
A compile receipt and a backtest result are real artifacts, and the assistant can point at them. If it cannot, that is your signal the claim is unverified.
Asking whether the edge is real
A backtest describes one window. Whether that description generalises is a separate question, and the connector can answer it with the same eight checks the Edge Report shows in the app.
A deep backtest already carries five of them: the out-of-sample check, walk-forward, track record, the trade-order test, and which market conditions produced the return. The other three — random entries, shuffled market, and parameter sensitivity — are deferred, because each re-runs the whole money model many times over. Ask for them:
- "Is this edge real, or just this window?"
- "Run the Edge Check on that backtest."
- "Would randomly timed trades have done as well?"
- "Does it still work one setting either side of that?"
They cost none of your daily backtest runs — nothing is re-simulated. A sweep can take a while, so the assistant starts it and checks back rather than blocking.
Unrun is not the same as passed
Until you ask, those three reads have no answer, and the connector reports them as not run rather than as clear. An assistant that calls a strategy validated without them has skipped the part that decides.
What it will refuse
Deployment, live trading, exchange credentials, and anything that moves money. Those tools do not exist in the connector, so the answer is not a policy refusal — there is simply no way to do it from here. Deploy from the Spera app when you are ready.
Next
- When it doesn't connect
- Backtesting — what the numbers mean