Tutorial · MCP

Grade a strategy inside Manus

Manus will happily write you a trading strategy. We graded its best one: F. Here's how to make Manus check its own homework — connect Tessen's grader as a custom MCP connector and every strategy it writes can be graded before it touches money, without leaving the conversation.

Step 1 — get a free API key

Discovery and the schema tool are open, but grading runs a real 6-year backtest, so the grading tools authenticate with your own key. Create one free at tessen.ai/studio/keys — it looks like tsn_live_….

Step 2 — add the connector in Manus

In Manus: Settings → Connectors → + Add Connectors → Custom MCP tab. The fastest path is Import by JSON — paste this and replace the placeholder with your key:

{
  "mcpServers": {
    "tessen": {
      "url": "https://tessen.ai/api/mcp",
      "headers": {
        "X-API-Key": "tsn_live_YOUR_KEY_HERE"
      }
    }
  }
}

Using Direct Configuration instead: name it tessen, set the server URL to https://tessen.ai/api/mcp, choose the HTTP transport (pick "Streamable HTTP" where offered — the server speaks MCP Streamable HTTP, protocol 2025-03-26), and add an X-API-Key header with your key if the form supports headers. No OAuth involved; the key is the whole handshake.

Step 3 — the tools Manus gets

get_agentspec_referencefree, no key

Returns the exact AgentSpec JSON schema, the indicator list, and the grade-report shape. Free, no key — the agent should call this first so it constructs a valid spec.

grade_strategyneeds API key

Backtests and grades an AgentSpec on 6 years of real data: 60/40 chronological train/OOS split, fees modeled, five hard gates. Returns the full report and a permanent public verify URL.

grade_trade_recordneeds API key

Grades your own trade history (30+ trades) instead of a backtest — same gates, same split. Marked “Checked” rather than “Verified” since the data is self-supplied.

Step 4 — use it in a conversation

Prompts that work well:

  • "Design your best mechanical BTC/ETH strategy, then use the tessen connector to grade it. Fetch the AgentSpec reference first, express the strategy as a valid spec, call grade_strategy, and tell me honestly whether it passed the five gates."
  • "Here are the rules of a strategy I'm considering: … Convert it to an AgentSpec and grade it with tessen before we discuss position sizing."
  • "Iterate: propose a variant, grade it, and stop when a variant passes all gates or after five attempts. Report every grade, not just the best one."

Every grade_strategycall returns a permanent public verify URL — share it, and anyone can confirm the grade wasn't cherry-picked. That last prompt is worth taking seriously: most variants will fail, and an agent that reports only its best backtest is doing exactly the thing that loses people money.

Troubleshooting

  • 401 from grading tools— the key isn't reaching the server. Prefer the JSON import with the headersblock; some connector forms don't forward custom headers from Direct Configuration.
  • Connector won't initialize — make sure the transport is the HTTP/Streamable option, not stdio. The endpoint only accepts POSTed JSON-RPC.
  • "insufficient OOS trades" grade— that's honest, not broken: the spec fires too rarely for 30+ out-of-sample trades. Loosen the entry confluence or widen the universe and re-grade.

Why bother?

Because AI-generated strategies look institutional whether or not they have edge — our case study on Manus's own best strategy is the proof. Grading is free, the gates are published, and the verify link is permanent. Works the same from Claude, Cursor, or any MCP client — the connector config above is all there is.