Oraclaw

What Whatsonyourmind/oraclaw MCP server does

Whatsonyourmind/oraclaw MCP server gives an MCP-compatible agent access to numerical methods for decision support. The supplied catalog covers 19 algorithms and 28 MCP tools across optimization, probability, simulation, forecasting, graph analysis, scheduling, and risk assessment. Tool calls return structured results such as selected options, objective values, confidence intervals, assignments, paths, anomaly details, or model metrics.

Use it when the input can be represented as numeric observations, variables, graph edges, time-series values, probabilities, or explicit constraints. Examples include choosing an A/B test arm, allocating a budget under caps, estimating uncertainty, combining forecasts, finding a route, or matching tasks to time slots.

How it works

The agent selects an MCP tool and supplies the relevant data for that operation. Bandit tools choose among alternatives using UCB1, Thompson sampling, epsilon-greedy, or LinUCB. Continuous black-box objectives can be handled with CMA-ES, while discrete or mixed search spaces can use a genetic algorithm. Linear, mixed-integer, and quadratic programs are handled by the HiGHS-based constraint solver.

Other tools operate on supplied datasets rather than external services. Forecasting accepts an evenly spaced numeric series and uses ARIMA or Holt-Winters. Risk analysis calculates VaR and Conditional VaR from historical returns and portfolio weights. Graph analysis accepts a directed weighted graph and can calculate PageRank, communities, bottlenecks, and an optional critical path. Monte Carlo simulation samples from one of six supported distributions and returns summary statistics and a histogram.

Some operations are stochastic, including Monte Carlo sampling and evolutionary optimization, so repeated calls may produce different results. Forecasting and anomaly detection also have minimum data requirements described by their respective tool behavior.

Setup and configuration

{
  "mcpServers": {
    "oraclaw": {
      "command": "npx",
      "args": ["-y", "@oraclaw/mcp-server"]
    }
  }
}

Whatsonyourmind/oraclaw MCP server can therefore be started through npx without a separate hosted endpoint in the client configuration. The README demonstrates this setup with Claude Desktop. Free tools do not require an API key according to the catalog. Premium tools accept ORACLAW_API_KEY; the tool descriptions also identify per-call x402 payment as an alternative for those operations.

The repository additionally documents a hosted REST API and separate npm SDK packages, but those are different access methods from the MCP server setup above.

Tools and capabilities

• Select an arm from historical reward data or context features with optimize_bandit and optimize_contextual.
• Search continuous, discrete, mixed, or multi-objective spaces with optimize_cmaes and optimize_evolve.
• Solve explicit allocation constraints with solve_constraints, or match tasks to energy-rated time slots with solve_schedule.
• Forecast series, detect outliers, simulate distributions, and compare base-case scenarios.
• Combine predictions, update Bayesian beliefs, score calibration, and measure source convergence.
• Route through weighted graphs with plan_pathfind, or inspect graph structure with analyze_graph.

Limitations and notes

Premium access applies to CMA-ES optimization, constraint solving, graph analysis, risk analysis, forecasting, and anomaly detection. A key or x402 payment is needed for those tools. The free REST tier is documented as allowing 25 calls per day per IP, but that limit is stated for the API rather than explicitly for local MCP execution.

Tool selection depends on the shape of the problem. Use optimize_bandit without per-call features, optimize_contextual when context changes between calls, and plan_pathfind for a single route rather than graph centrality. predict_forecast requires at least 20 observations for ARIMA and at least twice the seasonal length for Holt-Winters. Monte Carlo calls are capped at 2,000 iterations. Results should be checked against the supplied data and objective definition, particularly for stochastic algorithms and model-based forecasts.