indicato-mcp

An MCP server that gives an AI assistant the 89 indicators in indicato.

The assistant asks for an RSI; this computes one. It does not guess at the arithmetic, and it does not hallucinate a number that looks about right.

Install

claude mcp add indicato -- npx -y indicato-mcp

Or, for any client that reads a JSON config:

{
  "mcpServers": {
    "indicato": {
      "command": "npx",
      "args": ["-y", "indicato-mcp"]
    }
  }
}

Node 18 or newer. Nothing else to configure.

Its only dependency is indicato. The MCP protocol is spoken directly — newline-delimited JSON-RPC on stdin and stdout — rather than through the reference SDK, which declares two web frameworks, CORS, OAuth PKCE and a JOSE implementation to carry transports this server does not use. That was 88 packages behind a process that opens no socket and reads your files.

Where the data comes from

Nowhere. This server has no network access and no market data of its own — the same rule the library follows. Bars arrive one of two ways:

Inline, for a few hundred bars:

{ "indicator": "rsi", "bars": [{"o":100,"h":102,"l":99,"c":101,"v":10000}, ...] }

A plain array of numbers is accepted too, and treated as closes.

From a file on the machine running the server, for anything larger:

{ "indicator": "rsi", "file": "/data/RELIANCE.csv", "params": { "n": 21 } }

CSV headers are matched case-insensitively against open/high/low/close/volume and the usual variants — Adj Close, close_price, LTP, and so on. The columns actually used come back in the response, so a wrong guess is visible rather than silent. If the rows are newest-first and carry dates, they are reversed and you are told.

Tools

| tool | what it does | |---|---| | list_indicators | All 89, with parameters and defaults. Start here. | | describe_indicator | Parameters and return shape for one. | | compute | One indicator over your bars. | | compute_many | Several indicators, one read of the data. |

compute returns the trailing values plus a summary — warmup index, first, last, min, max, mean — rather than the whole series. Twenty years of daily bars through bollinger is three arrays of five thousand floats, which no model needs to read to answer a question about the last month. Pass full: true when you actually want everything, and tail: 60 when you want more than the default ten.

{
  "indicator": "macd",
  "values": { "line": [4.21, 4.33, ...], "signal": [...], "hist": [...] },
  "summary": { "line": { "warmup": 33, "last": 4.33, "min": -8.1, "max": 11.4, ... } },
  "source": { "file": "/data/RELIANCE.csv", "rows": 2480, "columns": { "c": "Close" } },
  "showing": "last 10 of 2480"
}

Two indicators need a benchmark

beta and correlation relate one instrument to another. Pass a benchmark series of the same length as the bars, or a benchmarkFile:

{ "indicator": "beta", "file": "/data/RELIANCE.csv", "benchmarkFile": "/data/NIFTY50.csv" }

Lengths must match bar for bar. A mismatch is refused rather than quietly truncated — a beta computed against a misaligned index is worse than no beta.

Mistakes it will not let you make

A wrong parameter name is an error, not a shrug:

rsi has no parameter period. It takes: n = 14.

A wrong indicator name suggests the right one:

no indicator named bollingerbands. Did you mean: bollinger, bbWidth, bbPercent?

Both are addressed to the model, which reads the message and corrects itself.

What it can reach

The file system, as you, and nothing else. It opens no network connection.

file is a path, and the thing choosing that path is a language model, which can be talked into things by what it reads. So a file that fails to parse is not quoted back into the error message — only fragments that look like column names survive, which Adj Close does and a line of a private key does not. Files are capped at 64 MB, and window lengths are bounded before anything is allocated.

Point it at a directory of exports and it is useful. It has no more reach than that.

Licence

MIT.