Overview

Semilattice MCP is a remote Model Context Protocol (MCP) server that predicts how defined audiences will respond to questions or pieces of content. It exposes audience population models and prediction tools over Streamable HTTP with OAuth, enabling language models, agents, and developer tools to query likely answers, probabilities, and aggregated statistics for specific audience segments.

This server is useful for A/B testing, content validation, personalization, feature experimentation, and social-science style research. Instead of running new surveys or relying solely on heuristics, agents can query Semilattice MCP to estimate how target audiences (e.g., “developers”, “marketing managers”, “Gen Z users”) would respond to a prompt or choice under test.

Repository and docs

GitHub: https://github.com/semilattice-research/mcp
Full docs: https://docs.semilattice.ai/mcp/introduction
MCP base URL: https://mcp.semilattice.ai

Features

Audience population models: curated or inferred models that represent demographic, professional, or psychographic segments.
Prediction API: estimate answers, distributions, and confidence scores for audience segments.
Streamable HTTP transport: supports streaming responses when available for large results or stepwise interactions.
OAuth-based authentication: integrates with provider-based sign-in flows for secure access.
Easy integrations: pre-built examples and configurations for popular MCP-capable clients and agent frameworks.
Designed for agent/tool workflows: methods to list available audience models, fetch metadata, and request predictions programmatically.

Installation / Configuration

Semilattice MCP runs as a remote service — you don’t need to host it locally. To connect, point your MCP-capable client to the MCP URL and authenticate with OAuth.

Core URL:

https://mcp.semilattice.ai

Example client configurations

Cursor (mcp.json)

{
  "mcpServers": {
    "semilattice": {
      "url": "https://mcp.semilattice.ai",
      "headers": {}
    }
  }
}

Cline (plugin settings snippet)

{
  "mcpServers": {
    "semilattice": {
      "type": "streamableHttp",
      "url": "https://mcp.semilattice.ai"
    }
  }
}

Gemini CLI (settings.json)

{
  "mcpServers": {
    "semilattice": {
      "httpUrl": "https://mcp.semilattice.ai"
    }
  }
}

Claude Code (CLI add command)

claude mcp add --transport http semilattice https://mcp.semilattice.ai

ChatGPT, Claude desktop/web, and other GUI clients typically provide an “Add custom MCP / Connect” option; supply the MCP URL above and follow the OAuth sign-in flow.

Available Tools / Available Resources

Semilattice exposes a small set of high-level tools/operations that agents can call. Names below are representative — consult the API docs for exact operation names and parameters.

listAudienceModels
- Returns available audience models, IDs, and brief metadata.
getAudienceDetails
- Fetch demographic distributions, model provenance, and sample sizes.
predictForAudience
- Given a prompt or question and an audience ID, returns predicted answer(s), choice probabilities, and summary statistics.
batchPredict
- Submit multiple prompts/questions and receive streamed or batched predictions.

Example (pseudo-request)

POST /mcp/predictForAudience
Authorization: Bearer <token>
{
  "audienceId": "developers_v1",
  "prompt": "Do you expect AGI within 10 years?",
  "options": ["Yes", "No", "Unsure"]
}

Example (pseudo-response)

{
  "audienceId": "developers_v1",
  "predictions": {
    "Yes": 0.32,
    "No": 0.45,
    "Unsure": 0.23
  },
  "metadata": {
    "modelVersion": "2026-01",
    "sampleSizeEstimate": 12000
  }
}

Use Cases

A/B testing content: Before launching, ask “Which headline will resonate more with marketing managers?” and use predicted choice probabilities to prioritize variants.
Feature personalization: Predict which UI affordances different user segments prefer and toggle defaults per audience.
Campaign targeting: Validate message framing by estimating how political, professional, or age-based cohorts respond.
Research prototyping: Rapidly obtain hypothesis-driven audience response estimates for exploratory social-science work without running immediate surveys.
Assistants and agents: LLMs can call the MCP during multi-step reasoning to ground recommendations in audience-aware predictions.

Concrete example

Scenario: A product team wants to choose between two onboarding flows A and B for “early-career developers”.
1. listAudienceModels -> find audience ID “developers_earlycareer_v2”
2. predictForAudience for prompt “Which onboarding would you prefer?” with options [“A”, “B”]
3. Use predicted probabilities to choose the higher-performing flow before running a live experiment.

Best Practices & Notes

Interpret predictions as model-based estimates, not exact ground-truth. Use them to prioritize experiments and hypotheses, not as a final substitute for live A/B test data.
Prefer aggregated metrics (probabilities, distributions, confidence intervals) over single-sample outputs.
Document which audience model/version you used; model drift or updates change predictions over time.
Use streaming where supported for large batches or when integrating into LLM agents that expect partial results.

Troubleshooting & Security

Authentication: Ensure OAuth flow completes and tokens are stored according to your client’s recommended practice.
Rate limits: Respect client and server rate limits; batch requests when possible.
Privacy: Avoid sending personally identifiable information to the MCP. Predictions are population-level; inputs should be anonymized where possible.
If you see “disconnected” in a client UI, re-run the client’s MCP authentication step or consult the client docs for re-auth steps.

For full API reference, integration examples, and model documentation, see the docs: https://docs.semilattice.ai/mcp/introduction and the repository on GitHub: https://github.com/semilattice-research/mcp.

Semilattice MCP Server: Audience Prediction for A/B Tests