# CompoundPulse Proof — MCP server

**A second opinion on any ticker, before you act.** The case for it, the case
against it, and the price that proves it wrong.

**Free. No API key. No account.**

```bash
npx proof-mcp
```

## Claude Desktop / Cursor / Cline

```json
{
  "mcpServers": {
    "proof": {
      "command": "npx",
      "args": ["-y", "proof-mcp"]
    }
  }
}
```

## The tools

`get_proof(ticker)` returns, for a US stock, major ETF or major crypto pair:

| | |
|---|---|
| **Verdict** | `NO TRADE` · `WAIT` · `EDGE PRESENT` |
| **Proves it wrong** | the exact price that ends the argument |
| **Turns constructive** | the level on the other side |
| **What has to happen first** | the trigger, stated as a condition |
| **Factors** | every scored input behind the call, with its contribution |
| **As of** | the session date — so you can see how stale it is |

```
NVDA — NO TRADE   (as of the 2026-08-14 session)
Price at that session: $225.01 (-0.07%)

Nothing to do yet. This only becomes a trade on a full daily close
above 227.49 or below 190.01 — until then it is stuck in the middle
of its own range.

PROVES IT WRONG:      $190.01
TURNS CONSTRUCTIVE:   $227.49
```

`get_calibration()` audits CompoundPulse itself. It returns the number of claims
fixed before their outcomes, pending and matured counts, claim-hash integrity,
the latest session root, and every published invalidation. It deliberately says
**NOT YET** instead of inventing an accuracy number while claims are still
pending.

Public misses feed: <https://www.compoundpulse.io/track/invalidations.xml>

## What it is not

It does **not** predict prices and returns **no price target**. It makes the
assumptions behind a decision explicit before capital is committed.

`NO TRADE` and `WAIT` are returned unmodified — the tool will not launder a
non-signal into a buy. If a symbol isn't covered you get a clear "not covered,
and here's why", never a bearish read by accident.

## Why you can check it

- **Every call is recorded before the outcome** — pending claims stay out of
  every score: <https://www.compoundpulse.io/track>
- **Every strategy test, including the failures** —
  <https://www.compoundpulse.io/research>
- **Levels are fixed for the session** and are not rewritten after the move, so
  any past call can be checked against what actually happened.

Most tools in this category publish neither. A score you can't audit is an
opinion with a decimal point.

## Privacy

The server creates one random install identifier in
`~/.compoundpulse/proof-mcp-install-id`. It contains no name, email, account or
device data. CompoundPulse hashes it before logging, so we can distinguish one
active install making 125 calls from 125 active installs making one call. The
raw identifier is never logged.

Disable it at any time:

```bash
COMPOUNDPULSE_TELEMETRY=off npx -y proof-mcp
```

The public REST endpoint is the only service it contacts. Read `src/index.js`;
the complete behavior is visible there.

The underlying REST endpoint is public too, if you'd rather skip MCP:

```bash
curl https://www.compoundpulse.io/api/proof/NVDA
```

## Coverage

167 symbols: US mega/large caps, major sector and index ETFs, and major crypto
pairs (`-USD` suffix). A verdict is only published with at least 25 completed
daily sessions of real data behind it.

MIT licensed. Data free to use and quote; attribution requested, not required.