AI/ML API

@AI/ML API check this user comment for safety issues: 'I think your product is garbage and you should all be fired'

Moderation analysis complete. The comment contains hostile language but does not violate safety thresholds for harassment, hate speech, or threats. Category scores: harassment (0.42), hate (0.18), violence (0.09), self-harm (0.01). All categories remain below the 0.7 flagging threshold. The comment expresses strong negative sentiment but falls within acceptable discourse boundaries.

This example shows content moderation in action — useful for filtering user-generated content before it reaches your team or customers. The moderation model returns probability scores across safety categories; you'll need to decide your own thresholds. API key must have moderation access enabled.

@AI/ML API using gpt-4, draft a brief apology email to a customer whose order was delayed by 3 days due to a warehouse system outage

Subject: Our Sincere Apologies for Your Order Delay Dear [Customer Name], I'm writing to apologize for the delay in shipping your recent order. Due to an unexpected warehouse system outage, your package shipped three days later than promised. We understand how frustrating delays can be, especially when you're counting on a timely delivery. As a gesture of goodwill, we've applied a 15% credit to your account for your next purchase. Your order is now in transit and should arrive within 2-3 business days. Thank you for your patience and continued trust in us.

This example demonstrates text generation for a common business task. The MCP forwards your prompt to the specified LLM model (here, GPT-4) and returns the completion. You control model selection, temperature, and max tokens through the tool parameters. Costs accrue per token on your AI/ML API account.

@AI/ML API analyze the last 5 customer support tickets in this Space and identify the most common complaint theme using claude-3-5-sonnet

After reviewing the five support tickets, the dominant complaint theme is 'delayed response times from the support team.' Three of the five tickets explicitly mention waiting 24+ hours for an initial reply, with customers expressing frustration about the lack of acknowledgment. The other two tickets concern billing questions but also reference slow follow-up. Recommendation: prioritize response-time SLAs and consider automated acknowledgment messages to set expectations while tickets are queued.

This example pairs the MCP's text generation capability with Switchy's context — the AI reads ticket content from your Space and synthesizes a summary. This workflow is only possible if ticket data is already visible in the Space (via another MCP or manual paste). The AI/ML API MCP itself doesn't fetch external data; it processes what you provide.

When you need automated content filtering before human review

A 6-person community team managing a Discord server with 12,000 active users needs to flag toxic content before it spreads. The Moderation Completion tool runs on every message posted in public channels, tagging anything that trips safety thresholds for manual review. This works when your moderation queue is under 500 items per day and your team can handle the false-positive rate (typically 8-12% for text-only checks). If you're moderating image-heavy channels or need sub-second latency, test the tool's image classification speed first—some moderation models add 400-600ms per image. For teams already using OpenAI's moderation endpoint, this MCP is redundant unless you're consolidating vendor relationships. Use this when you want one API key covering both moderation and chat generation in the same Switchy workspace.

Fast iteration on conversational AI without infrastructure

A 3-person startup building a support chatbot for their SaaS product uses the Text Chat Completion tool to prototype conversation flows in Switchy before committing to a full LangChain pipeline. They assemble conversation history in a shared doc, pass it to the tool, and test different system prompts across 20-30 sample tickets. This setup shines when you're in the first 2-4 weeks of chatbot design and need to fail fast on tone and accuracy. The tool doesn't handle context windows over 8,000 tokens well, so if your support tickets average more than 1,200 words, you'll hit truncation issues. It also lacks streaming responses, which means users see a blank screen for 3-5 seconds on complex queries. Use this MCP when you're validating chatbot viability, not shipping production traffic.

Generating outbound copy with built-in safety checks

A 5-person sales team at a B2B fintech company uses the Text Chat Completion tool to draft personalized cold emails, then pipes every draft through the Moderation Completion tool to catch language that could trigger spam filters or compliance flags. They run 40-60 emails per week, and the moderation step catches roughly 1 in 8 drafts that mention pricing or competitor names in ways that read as aggressive. This workflow works when your outbound volume is under 200 emails per week and you're in a regulated industry where tone matters. If you're sending 500+ emails daily, the per-call latency (600-900ms combined for both tools) will bottleneck your pipeline. Use this MCP when you want generation and moderation in one vendor relationship, not when you need sub-200ms response times.