Anthropic MCP Integration

Weldable's Anthropic MCP integration turns Claude into a callable step inside your automated workflows. Instead of using Claude only as the agent running a workflow, you can call Claude from within a workflow to generate text, analyze data, classify inputs, or make decisions at any point in the pipeline. One action, "Ask Claude," and it handles the rest.

This is a different pattern from most integrations. Slack sends messages. Google Sheets reads rows. The Anthropic integration thinks. It takes unstructured input, reasons about it, and returns structured output that downstream steps can act on.

Use cases

Inbound lead qualification

A workflow reads new form submissions from Google Sheets, passes each one to Claude with a system prompt defining your ideal customer profile, and gets back a JSON object with a score, reasoning, and recommended next step. High-scoring leads get a Slack notification to your sales channel. Low scorers get an automated follow-up email via Gmail.

Daily content briefing

Every morning, a workflow pulls trending topics from an RSS feed, sends them to Claude with instructions to identify the three most relevant to your industry, and drafts a short briefing for each. The summaries post to a Slack channel before your team's standup. Claude handles the editorial judgment that a simple keyword filter cannot.

Support ticket routing

When a new support email arrives, Claude reads the message and classifies it by urgency, product area, and customer sentiment. It returns structured JSON that the workflow uses to route the ticket: billing issues go to finance, bugs go to engineering with a severity tag, and feature requests land in a product backlog sheet.

Code review summaries

A workflow triggers on new GitHub pull requests. It fetches the diff, sends it to Claude with a system prompt focused on security and performance patterns, and posts a summary comment on the PR. The model flags potential issues and explains why they matter, giving reviewers a head start.

Multi-language content adaptation

Your marketing team writes blog posts in English. A workflow takes the published content, sends it to Claude with instructions to adapt (not just translate) for three target markets, adjusting cultural references and tone. Each adaptation writes to a separate Google Doc, ready for native-speaker review.

How it works

The "Ask Claude" action sends a prompt to the Anthropic API and returns the response as data your workflow can use in subsequent steps. You write the prompt in natural language, just like you would in a chat, but you can inject variables from earlier steps.

Under the hood, each call goes through Weldable's worker with your Anthropic API key. You pick the model: Haiku for fast, cheap classification tasks; Sonnet for the sweet spot of speed and quality; Opus when accuracy matters more than latency. An optional system prompt sets the persona or constraints for that specific step.

The schema parameter is where this gets powerful. Pass a JSON schema and Claude's response will conform to that exact structure. This means the next step in your workflow can reliably access response.score or response.category without parsing free text. No regex, no brittle string matching.

Tips

Use system prompts to scope the task. A generic "analyze this" prompt produces generic results. A system prompt like "You are a senior security engineer reviewing code diffs. Flag only issues with real exploit potential. Ignore style." produces focused, actionable output.

Match the model to the job. Haiku costs roughly 1/10th of Opus and responds in a fraction of the time. For binary classification, sentiment analysis, or simple extraction, Haiku is the right pick. Save Sonnet and Opus for tasks that need nuanced reasoning or long-context understanding.

Always use the schema parameter for structured data. When a downstream step needs specific fields, define a JSON schema instead of asking Claude to "respond in JSON format" in the prompt. The schema parameter guarantees conformance. The prompt instruction does not.

Keep prompts short and specific. Long, over-engineered prompts often perform worse than concise ones with clear constraints. State what you want, give one example if the format is unusual, and stop. Claude responds well to direct instructions.

Put large inputs above your instructions. If you're passing a document or dataset for analysis, place it at the top of the prompt with your question below. This matches how Claude processes context and produces better results, especially with long inputs.

Chain multiple Claude calls for complex reasoning. A single prompt that asks Claude to "read this email, classify it, draft a response, and translate it" will underperform compared to four separate steps, each with a focused prompt and its own schema. Workflow steps are cheap. Bad outputs are expensive.