ta-mcp

ta-mcp

MCP server that lets AI agents run TA-Lib technical-analysis calculations on price data they supply.

It is a stateless calculator: the agent sends Series, and the server returns Indicator values. It never fetches market data and keeps nothing between calls. The catalogue covers 174 Indicators across TA-Lib's eight analysis groups. The Math Operators and Math Transform groups are left out.

Run

Requires Python 3.12 or later. The ta-lib wheels bundle the TA-Lib C library. ta-mcp isn't on PyPI yet, so uvx installs it from GitHub:

uvx --from git+https://github.com/T-bytes/ta-mcp ta-mcp                  # stdio transport
uvx --from git+https://github.com/T-bytes/ta-mcp ta-mcp --max-bars 5000  # lower the per-call Bar Cap
uvx --from git+https://github.com/T-bytes/ta-mcp ta-mcp --max-steps 20   # lower the per-call Step Cap

Setting	Flag	Environment	Default
Bar Cap (most Bars per call)	--max-bars	TA_MCP_MAX_BARS	10000
Step Cap (most Steps per call)	--max-steps	TA_MCP_MAX_STEPS	50

A flag takes precedence over its environment variable.

Claude Code

claude mcp add ta-mcp -- uvx --from git+https://github.com/T-bytes/ta-mcp ta-mcp

Claude Desktop

{
  "mcpServers": {
    "ta-mcp": {
      "command": "uvx",
      "args": ["--from", "git+https://github.com/T-bytes/ta-mcp", "ta-mcp"]
    }
  }
}

Tools

Tool	Purpose
list_indicators(group?)	Names, display names and Groups.
describe_indicator(names)	For each name: the Default Series it reads, Parameters (type, default, help, MA Type choices), Outputs, lookback and caveats.
compute_indicators(inputs, steps, timestamps?, tail?, precision?)	Runs each Step (id, name, params?, input_mapping?, include?) over one shared set of Series. For each Step id: Outputs aligned Bar-for-Bar with the input, null for the lookback period, and the lookback count.

Example

A compute_indicators call:

{
  "inputs": {"close": [44.3, 44.1, 44.2, 43.6, 44.3]},
  "steps": [{"id": "bb", "name": "BBANDS", "params": {"timeperiod": 3, "matype": "EMA"}}],
  "timestamps": ["2026-09-21", "2026-09-22", "2026-09-23", "2026-09-24", "2026-09-25"],
  "tail": 2,
  "precision": 2
}

returns results keyed by Step id. timestamps comes back only if you sent it, and tail trims it along with the Outputs:

{
  "results": {
    "bb": {
      "name": "BBANDS",
      "outputs": {
        "upperband": [44.42, 44.72],
        "middleband": [43.9, 44.1],
        "lowerband": [43.38, 43.48]
      },
      "lookback": 2
    }
  },
  "timestamps": ["2026-09-24", "2026-09-25"]
}

To chain Steps, point input_mapping at an earlier Step's Outputs. This call reads doji patterns off Heikin-Ashi candles instead of the raw OHLC:

{
  "inputs": {"open": [...], "high": [...], "low": [...], "close": [...]},
  "steps": [
    {"id": "ha", "name": "HA", "include": false},
    {"id": "doji", "name": "CDLDOJI", "input_mapping": {
      "open": "ha.haopen", "high": "ha.hahigh", "low": "ha.halow", "close": "ha.haclose"
    }}
  ]
}

Rules

• Series are columnar arrays, oldest Bar first, all the same length. The Default Series are named open, high, low, close and volume; MAVP also reads periods. describe_indicator lists the ones each Indicator reads. Series no Step reads are ignored.
• Input Mapping reads another Series in place of a Default Series, e.g. {"close": "obv.real"}. Use it to chain Indicators or to feed Heikin-Ashi candles into pattern Indicators. CORREL and BETA read series_a and series_b.
• Chaining: a Step reads an earlier Step's Output as <id>.<output>, at full precision and over the full history whatever tail and precision say. Set include: false on Steps you only chain. Step ids can't contain ., and a Step's Output can't replace a Series of the same name.
• Leading nulls are allowed, for example the lookback of a chained Output. A null after the first real value is rejected.
• MA Types are given by TA-Lib name (SMA, EMA, WMA, DEMA, TEMA, TRIMA, KAMA, MAMA, T3, HMA, ZLEMA, RMA).
• Candlestick Signals are integers: positive is bullish, negative bearish, 0 means no pattern. A few patterns use 200/-200 to mark a confirmed signal.
• Unstable Period: for some Indicators (EMA, RSI, ATR and others), values just after the lookback depend on how much history you send. Send several times the timeperiod in Bars. describe_indicator flags these Indicators.

Agent skill

skills/ta-analysis/ is an agent skill for using these tools well: preparing inputs, parameter conventions and platform parity, history sizing, ready-to-send recipes, and turning Outputs into signals without look-ahead bias. Copy it into ~/.claude/skills/ (or a project's .claude/skills/) to use it with Claude Code.

Design decisions are recorded in docs/adr/, and the domain vocabulary in CONTEXT.md.

Develop

uv sync
uv run pytest

License

BSD-2-Clause, the same as TA-Lib. See LICENSE.