OraClaw MCP Server

17 decision-intelligence tools your AI agent can call. Sub-25ms. 6 free without a key.

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

Drop that into your Claude Desktop / Cursor / Cline MCP config and your agent can immediately:

"I have 3 email subject line variants with these click rates. Which should I send next?"

The agent calls optimize_bandit → gets a statistically optimal arm in 0.01ms:

{ "selected": { "id": "C", "name": "Option C" },
  "score": 1.876, "exploitation": 0.9, "exploration": 0.976, "regret": 0.1 }

No LLM math, no hallucinated formulas, no setup beyond the snippet above.

All 17 tools

| Free (no key) | What it does | |---|---| | optimize_bandit | UCB1 / Thompson / ε-greedy arm selection | | optimize_contextual | LinUCB contextual bandit (per-situation choice) | | optimize_evolve | Genetic algorithm (discrete + multi-objective) | | solve_schedule | Task → time-slot assignment with energy matching | | score_convergence | Multi-source probability consensus (Hellinger) | | score_calibration | Brier + log score for forecaster accuracy | | predict_bayesian | Beta posterior update from weighted evidence | | predict_ensemble | Multi-model consensus + uncertainty decomposition | | plan_pathfind | A* + Yen's k-shortest paths | | simulate_montecarlo | Single-factor Monte Carlo (6 distributions) | | simulate_scenario | What-if comparison + sensitivity ranking |

| Premium (need API key) | What it does | |---|---| | optimize_cmaes | CMA-ES continuous black-box optimization | | solve_constraints | LP / MIP / QP solver (HiGHS, provably optimal) | | analyze_graph | PageRank + Louvain + critical path + bottlenecks | | analyze_risk | Portfolio VaR / CVaR with correlation | | predict_forecast | ARIMA / Holt-Winters time series | | detect_anomaly | Z-score / IQR outlier detection |

Every tool ships with an explicit inputSchema, outputSchema, and MCP behavior annotations (readOnlyHint, destructiveHint, idempotentHint, openWorldHint) so your agent knows exactly what it gets back.

Get a key for premium tools

oraclaw signup → — one email field, instant key, no card needed. 1,000 calls/day on pay-per-call ($0.005/call). Upgrade to Starter $9/mo for 50k/month.

Then add ORACLAW_API_KEY to your MCP env:

{
  "mcpServers": {
    "oraclaw": {
      "command": "npx",
      "args": ["-y", "@oraclaw/mcp-server"],
      "env": { "ORACLAW_API_KEY": "your-key-here" }
    }
  }
}

Use the API directly (no MCP)

curl -X POST https://oraclaw-api.onrender.com/api/v1/optimize/bandit \
  -H "Content-Type: application/json" \
  -d '{
    "arms": [
      {"id":"A", "name":"Option A", "pulls":10, "totalReward":7},
      {"id":"B", "name":"Option B", "pulls":10, "totalReward":5}
    ],
    "algorithm": "ucb1"
  }'

Free tier: 25 calls/day per IP, no key required.

Why agents need this

LLMs can't reliably do: bandit selection, LP optimization, time-series forecasting, graph centrality, Monte Carlo sampling, or anomaly detection. They confabulate the math. OraClaw is the deterministic substrate underneath your agent — every answer is a real algorithm, returns structured JSON, and runs in under 25ms.

• Source: github.com/Whatsonyourmind/oraclaw
• API docs: oraclaw-api.onrender.com
• Pricing: web-olive-one-89.vercel.app/pricing
• License: MIT