Give your AI assistant the denominator
Ask any chatbot "what's the best indicator for Bitcoin?" and it answers from forum folklore and indicator marketing — a corpus made almost entirely of winning backtests, because nobody publishes the losing ones. This MCP server hands the model the losing ones. 660,005 backtests across 903 assets, out-of-sample scored, and the one number the category never prints: an indicator beat buy-and-hold 20.1% of the time.
Connect it
One command, nothing to paste:
claude mcp add --transport http \ indicatoredge https://indicatoredge.io/mcp/v1
Add to your MCP config file:
{
"mcpServers": {
"indicatoredge": {
"url": "https://indicatoredge.io/mcp/v1"
}
}
}Settings → Connectors → Add custom connector. Paste the URL, leave auth empty:
https://indicatoredge.io/mcp/v1
Streamable HTTP, JSON-RPC 2.0, stateless. No session id, no SSE required, no auth header:
POST https://indicatoredge.io/mcp/v1 Content-Type: application/json
Smoke-test it in one command
Paste this in a terminal. If it returns a percentage, the server is up and your network can reach it — no client, no config, no account needed:
curl -s https://indicatoredge.io/mcp/v1 \
-H 'Content-Type: application/json' \
-d '{"jsonrpc":"2.0","id":1,"method":"tools/call",
"params":{"name":"ie_base_rate","arguments":{}}}'
Expected: base_rate_pct: 20.1. You can also just open /.well-known/mcp.json in a browser.
What this server can and cannot do
- Read published backtest results for 903 assets
- Return the base rate an indicator should be judged against
- Tell your assistant when the honest answer is "nothing worked"
- Place an order, cancel one, or size a position on a venue
- Hold an exchange credential — there is no field to put one in
- Move, hold, or custody funds of any kind
- Be prompt-injected into a trade, because no trade endpoint exists
Connecting an agent to anything that can trade deserves a careful read of what you are handing it. We removed the need for that read by removing the capability: this server has no write surface, so there is no key to leak, no allowance to revoke, and no worst case beyond a wrong number. Every tool is annotated readOnlyHint: true so your client can see that too.
Trading-data MCP servers are multiplying, and nearly all of them sell the same thing: live prices, funding rates, order flow — data about right now, usually metered per call. None of it answers the question that decides whether a trade is worth taking, which is does this signal work, and how often. That question needs a research archive, not a market feed. We have one, and handing it to an agent costs us nothing, so it is free.
What your assistant can call
How often a technical indicator actually beat buy-and-hold out-of-sample — 20.1% across 660,005 scored backtests. Pass an indicator key for its individual rate.
The empirically strongest indicator and timeframe for one ticker, with out-of-sample Sharpe, max drawdown and alpha vs buy-and-hold — or verdict NO_EDGE when nothing worked.
VALIDATED, MIXED or NOTHING, after a hurdle corrected for how many combinations were searched. NOTHING is the answer for 703 of 903 assets.
Search all 903 covered assets across stocks, ETFs, crypto, forex, commodities and indices.
All 382 indicators with how many assets each is the best fit for, average Sharpe, and its own beat-buy-and-hold rate.
The 10 setups that cleared out-of-sample validation — returned alongside the 703 failures, so the survivors read as a fraction, not a highlight reel.
Costs modelled, out-of-sample split, and an explicit list of what this dataset does not do.
What this server is and isn't
| Price | Free. No API key, no account, no wallet, no per-call micropayment. |
| What it can do | Read published backtest results. That is the whole surface. |
| What it cannot do | Place an order, hold a credential, touch an exchange, move a dollar. |
| Where the numbers come from | 660,005 backtests we ran and published, out-of-sample scored, standard textbook settings, ~0.08% per side in costs. |
| What it returns when there's no edge | NOTHING — for 703 of 903 assets. The negative answer is the product. |
Four ready-made asks
Is this indicator worth using? Audits one indicator against its measured base rate, and against the all-indicator average.
The honest answer for one ticker — including instructions not to substitute a weaker signal when the verdict is "nothing worked".
Reality-check a strategy you were sold, against the published record, before paying for it.
Where does anything actually work? Returns the survivors together with the failure count.
Your client will surface these as slash commands or prompt templates once connected.
Questions people ask
What is the IndicatorEdge MCP server?
A free, read-only Model Context Protocol server at https://indicatoredge.io/mcp/v1. It gives Claude, Cursor, or any MCP client direct access to 660,005 out-of-sample backtests across 903 assets, so the assistant can answer 'what's the best indicator for X' from measured evidence instead of from training-data folklore. It is read-only — it cannot trade.
How do I connect it to Claude?
Add {"mcpServers": {"indicatoredge": {"url": "https://indicatoredge.io/mcp/v1"}}} to your MCP client config. There is no API key step because there is no key — the data is public and so is the endpoint.
Does it cost anything per call?
No. Trading-data MCP servers commonly meter per call, often over crypto payment rails. We do not, because the value here is a published research dataset, and metering a research dataset just stops people checking their assumptions.
Can it place trades or connect to my exchange?
No — by design. There is no order endpoint, no credential storage, and no exchange integration anywhere in this server. Nothing here can lose you money or be tricked by a prompt injection into moving funds.
Why does it keep telling my assistant about a base rate?
Because a single winning backtest is meaningless without the denominator. Across our whole grid, an indicator beat buy-and-hold out-of-sample 20.1% of the time — roughly 1 in 5. Every tool here returns that context so the model reports a result as a candidate rather than a conclusion.
Prefer to read it yourself?
Get the weekly edge report
The best-performing indicator per asset, what changed this week, and the honest caveats — straight to your inbox.