Feature guides

Use Quantis Trade from your AI assistant

Quantis Trade runs an MCP server, so Claude, ChatGPT and other MCP clients can search our evidence, explain the numbers and run backtests for you, with a spending cap on every run.

On this page

What your assistant can do
Connect by signing in (Claude apps and ChatGPT)
Or create an API key
Use the key from Claude
Use the key from other clients
The tools
Costs
Limits
Privacy and security
Troubleshooting

What your assistant can do

The server lives at https://quantistrade.co.uk/mcp (MCP over Streamable HTTP). Your assistant can search the public evidence corpus, list the community strategy templates, explain any metric on a result card, estimate what a backtest costs, and, if you allow it, run backtests on your token balance through exactly the same Strategy Agent as the website: the same parser, prices, trial ledger and refunds.

There are two ways to connect, and both give the assistant the same tools, limits and activity log:

Sign in (OAuth), for the Claude apps and ChatGPT: you add the address above in the assistant, sign in to Quantis Trade and approve what it may do. Nothing to copy or paste.
A personal API key, for Claude Code, the Claude and OpenAI APIs, scripts, and any client that sends a request header.

Either way, nothing it does can trade, place orders, touch your paper portfolios, publish or share anything, change your settings or billing, delete anything, or read another user's data. A key or a connection doesn't work anywhere on the site except the MCP server.

Claude (claude.ai, Desktop and mobile): go to Customize > Connectors, choose Add custom connector, enter https://quantistrade.co.uk/mcp and add it. On Team and Enterprise plans an Owner adds the connector in Organization settings > Connectors, and each member then presses Connect.

ChatGPT: turn on Developer mode (Settings > Security and login), open the plugins page, press the plus button, give the connection a name and enter https://quantistrade.co.uk/mcp.

Claude Code can sign in too: add the server without a header (claude mcp add --transport http quantis https://quantistrade.co.uk/mcp) and run /mcp to sign in.

What happens next is the same in each:

The assistant sends you to quantistrade.co.uk. Sign in if you aren't already.
A Quantis Trade page shows which app is asking, the address your approval will be sent to, and each thing it wants to do. Check the address is the app you started from.
Read is always included. Run backtests (spends your tokens) is off unless you tick it: leave it off and the assistant can estimate costs and read results but can't spend anything.
Press Allow and you are sent back to the assistant, connected. Press Cancel and nothing is shared.

Every connection is listed under Connections > Connected apps with what it may do, when it was last used and its activity. Disconnect it there and it stops working on its next request. A connection an app stops using lapses on its own. To change what an app may do, disconnect it and connect again.

Or create an API key

Sign in and open Connections > API keys.
Name the key after where you'll use it (for example "Claude Code on my laptop"). Every key can read; tick Run backtests only if the assistant should be able to spend your tokens.
Choose when it expires (7, 30, 90, 365 days, or never), then create it.
Copy the key straight away. It starts qtk_ and is shown only once: Quantis Trade keeps only a fingerprint of it. Treat it like a password. If it leaks, revoke it on the Connections page and it stops working at once.

Use the key from Claude

Claude Code (the terminal and IDE app) sends the key as a request header:

claude mcp add --transport http quantis https://quantistrade.co.uk/mcp \
  --header "Authorization: Bearer YOUR_QUANTIS_KEY"

Or in a project's .mcp.json, reading the key from an environment variable:

{
  "mcpServers": {
    "quantis": {
      "type": "http",
      "url": "https://quantistrade.co.uk/mcp",
      "headers": { "Authorization": "Bearer ${QUANTIS_API_KEY}" }
    }
  }
}

Claude apps (claude.ai, Desktop and mobile): connect by signing in, as above. The connector dialog takes an API key only for organisations that have Anthropic's request headers option (in beta, set up by an organisation Owner, and shared by everyone in that organisation), and a shared key acts as one Quantis Trade account for all of them, so signing in is the better fit.

Claude API (Messages API MCP connector): pass the key as the server's authorization_token.

"mcp_servers": [
  { "type": "url", "url": "https://quantistrade.co.uk/mcp", "name": "quantis", "authorization_token": "YOUR_QUANTIS_KEY" }
],
"tools": [{ "type": "mcp_toolset", "mcp_server_name": "quantis" }]

Use the key from other clients

ChatGPT itself has no field for a key: connect it by signing in, as above. The OpenAI API's remote MCP tool accepts the key in its authorization field. Any other MCP client that can send an Authorization: Bearer header over Streamable HTTP works the same way as Claude Code.

The tools

Tool	Cost	What it does	Limits
`search_evidence` (any key or connected app)	Free	Searches the public evidence corpus of house backtests (fixed universes in the US, UK and EU; 5 to 20-year windows; the same costs). Returns matching runs with their metrics and links, counting refused and never-fired runs.	Stored runs only: it doesn't run anything and holds no user results.
`list_templates` (any key or connected app)	Free	The community strategy templates: each rule, its symbols, the exact prompt the Strategy Agent receives, and its house backtest when one is stored.	Doesn't run a template; pass its prompt to estimate_backtest_cost or run_backtest.
`explain_metric` (any key or connected app)	Free	The Education Centre's definition of a result-card metric (what it is, why it matters, the formula) with a link to the full article when one is published.	Definitions only; it doesn't compute a metric for a strategy.
`methodology` (any key or connected app)	Free	How a result card is calculated: costs, dividends and taxes; big falls when its history is reshuffled; could it be luck; where the gains came from; the evidence corpus; token pricing; concept articles.	Explains; doesn't compute.
`estimate_backtest_cost` (any key or connected app)	Free (nothing runs)	Reads a plain-English strategy exactly as the Strategy Agent would and returns what would run (symbols, period, the rules as the result card will state them), the price in tokens and your balance, or the clarification or refusal the agent would give.	Never charges.
`run_backtest` (needs the Run backtests scope)	Spends your tokens: 300 for every backtest, the same price as the website	Runs a strategy through the Strategy Agent: the same parser, prices, trial ledger and refunds as the website. Returns the verdict, key metrics, the exact rules, the cost basis and links to the full result card and its reproduce bundle, or a result id if it takes longer than the wait.	Needs max_tokens: a run above it, or above your balance, is refused before anything is charged. Daily data, at most 20 years. Goal prompts go to goal_search.
`get_backtest` (any key or connected app)	Free	Fetches one of your own results by id: status, verdict, metrics, rules, cost basis, tokens charged and refunded, links.	Your account's results only.
`goal_search` (needs the Run backtests scope)	Spends your tokens: 2,100 tokens for the default 24 candidates (as much as 7 tests, more than the free trial's 1,000 tokens)	Searches a fixed, documented list of candidate strategies for one meeting your goal, chosen on in-sample years and judged on held-out years, every candidate counted.	Needs max_tokens covering the quoted price. Usually takes minutes: it returns a result id for get_backtest.

Costs

Runs cost exactly what they cost on the website, and the assistant is told the price before it spends anything. Every backtest is 300 tokens, whatever its length (up to 20 years), number of symbols or cadence. A goal search costs 2,100 tokens for the default 24 candidates (as much as 7 tests, more than the free trial's 1,000 tokens). New accounts: verify your email to get 3 free tests (1,000 tokens).

Every run_backtest and goal_search call must carry max_tokens. If the run would cost more, or more than your balance, it is refused and nothing is charged.
Estimates, clarifications and refusals are free. A run whose rule never fired, or that the engine refuses, is refunded automatically.
A run waits up to 50 seconds by default (at most 110); if it takes longer the assistant gets a result id and fetches it with get_backtest. The run carries on either way.
Every tool call and what it charged is listed under the key's or the connected app's Activity on the Connections page, and every run appears in your Strategy Agent history and wallet ledger like a run from the website.

Limits

60 MCP requests a minute per key (and per connected app).
60 cost estimates, runs and goal searches an hour per account, of which at most 20 runs and goal searches.
Up to 10 active keys per account. Strategy descriptions up to 1,000 characters.
Backtests use daily data and at most 20 years, as on the website.

Privacy and security

We store a fingerprint (SHA-256) of each key, never the key itself, plus its name, first characters, scopes, expiry and when it was last used.
For an app you connected by signing in we store the app's name and address, the address your approval was sent to, what you allowed, and fingerprints (SHA-256) of its access and refresh tokens, never the tokens themselves. The app never sees your password.
For each key and each connected app we record an activity log: which tool was called, the outcome, tokens charged, the time, and the calling IP address and client name. The arguments of free tool calls aren't saved to your account or to the activity log. A run's prompt and result are saved to your account exactly as a run from the website is, and count in your private test ledger (the "Could it be luck?" line).
Your AI assistant's provider (Anthropic, OpenAI or another) receives whatever the tools return, under its own privacy terms. Tools return only public evidence, Education Centre content and your own results.
Keys can be revoked and apps disconnected at any time on the Connections page; they stop working at once. Deleting your account deletes its keys, connections and their activity.
Results are hypothetical backtests on historical data: not forecasts and not investment advice. See the security page and privacy policy.

Troubleshooting

401: the key is missing, mistyped, revoked or expired; the error says which. Check the header is Authorization: Bearer qtk_.... For a connected app, the error says whether the connection was disconnected or lapsed: connect again from the app.
"This connection can't continue" on the Quantis Trade page: the app's request couldn't be trusted (the page says why) and nothing was shared. Start the connection again from the app.
429: a rate limit; the response says when to try again.
No run_backtest tool: the key doesn't have the Run backtests scope (create one that does), or the app was connected without it (disconnect it, connect again and tick it).
A refusal with a clarification: the Strategy Agent needs more detail, exactly as on the website. Nothing was charged.

These pages describe how Quantis Trade works and what its tests show about the past. They are not forecasts and not advice.