RxInfer MCP

RxInfer MCP powers RxInfer Pro's developer tooling. It brings Bayesian decision support to your assistants and scores LLM responses with transparent uncertainty. This guide covers how to deploy it, plug it into different clients, and operate it safely.

Installation

RxInfer MCP is reachable at https://rxinfer-mcp-775335415984.europe-west4.run.app/mcp. The sections below cover the platforms that support remote MCP connectors today.

Claude.ai (web)

Available for Pro, Max, Team, and Enterprise plans.

1. Open Settings → Connectors.
2. Click Add custom connector.
3. Paste the RxInfer MCP URL.
4. Finish the flow and enable the connector via the Search & tools toggle inside a chat.

Claude Desktop

1. Open Claude Desktop → Settings.
2. Navigate to Developer → Model Context Protocol.
3. Click Edit Config and add:

{
  "mcpServers": {
    "rxinfer": {
      "url": "https://rxinfer-mcp-775335415984.europe-west4.run.app/mcp"
    }
  }
}

4. Restart the desktop client to load the configuration.

Cursor

1. Open Cursor Settings (⌘+, on macOS, Ctrl+, on Windows/Linux).
2. Search for Model Context Protocol.
3. Enable MCP and add the RxInfer MCP URL.
4. Restart Cursor.

ChatGPT (custom GPTs)

Official support is coming to the GPT Store. In the meantime you can wire RxInfer MCP via Actions:

1. Create or edit a GPT.
2. Go to Configure → Actions → Create new action.
3. Import from URL using the RxInfer MCP endpoint.
4. Test and publish your GPT.

Le Chat (Mistral)

Le Chat advertises MCP compatibility and we can establish a connection, but the product currently fails to pass the correct tool arguments. If you find a reliable workaround, please let us know or share it on the community Discord.

Contextual decisions

Use this workflow when you need action recommendations with quantified uncertainty.

You: “We’ve got churn history in CSV form. Learn the best retention move.”

Assistant: “Loading the data and choosing the right contextual decisions tool…” → calls learn_contextual_decisions → “Model trained; session sess_42 is ready.”

You: “For this at-risk customer, what should we do?”

Assistant: “Let me use the trained session…” → calls predict_contextual_action(sess_42, context) → “Recommend offer_credit. Expected uplift 0.62 with ±0.12 uncertainty. update_contextual_model is available if new outcomes arrive.”

Prepare the dataset

• Provide historical decisions with outcomes in CSV (or convertible) form.
• Include columns for relevant context features, the action taken, and the observed outcome.

customer_age,usage_days,support_tickets,action,revenue
25,30,2,standard,145.50
45,5,8,premium_support,89.20
35,90,1,discount,203.75
55,15,3,standard,112.00

Ask for recommendations

1. Share the dataset with your assistant.

2. Request probabilistic predictions. Example prompt:

   “Given this history, what action should we take for a 35-year-old customer with 60 days usage and 2 support tickets? Include confidence intervals and explain the risk trade-offs.”

Interpret the output

You’ll receive:

• Expected outcomes for each candidate action with confidence bounds.
• Probability of one action outperforming another.
• Downside-risk commentary (e.g. worst-case outcomes, high-variance warning).
• A recommendation that explicitly incorporates uncertainty, not just averages.

Example conversations

• Churn prevention strategy (Claude share)
• Lead qualification & prioritisation (Claude share)
• Onboarding new customers:
  • Claude link
  • ChatGPT link

LLM response evaluation

evaluate_llm_response grades model answers while exposing how certain the score is.

Parameters

• question (string, required) — The original question.
• answer (string, required) — The model response to evaluate.
• mode (optional) — One of quick, standard, or thorough. Defaults to standard.

Modes at a glance

Example interaction

You: “Did the assistant just hallucinate? Please review this answer: ‘Customer missed three invoices and dropped usage—flag churn risk?’”

Assistant: “Checking which tool applies…” → selects evaluate_llm_response → “Calling in thorough mode.”

Tool response: correct (0.72 ± 0.09). Alternatives: incorrect 0.11, unclear 0.09, others <0.05.

Interpreting the results

The evaluator returns:

• A truthfulness / reliability rating.
• Probability distributions showing alternative scores.
• Standard deviations and confidence intervals, signalling how trustworthy the evaluation itself is.

Platform support

Data guidelines

• For reliable results, aim for 50–100 observations per action with balanced coverage.
• The system can operate on as few as 20–30 observations; uncertainty intervals will simply widen.

Security

• RxInfer MCP processes data in-memory only; no persistent storage is kept.
• Prefer anonymised identifiers and aggregated metrics whenever possible.
• Use temporary file links that expire after upload, especially for customer-level datasets.
• For enterprise deployments, we support private hosting with full data sovereignty—get in touch via contact.

Integrations

Integrations

Connect RxInfer MCP outputs with the tools your GTM teams already use.

Salesforce

Route MCP insights straight into Sales Cloud workflows.

Talk to us

HubSpot

Trigger playbooks and score leads directly from RxInfer output.

Talk to us

Support

• Join the Lazy Dynamics Discord to collaborate with other builders.
• Submit questions, feedback, or deployment requests via contact.