Tendrl
Use Case
AI-Powered Pipelines
Add AI to any workflow without spinning up another service. Drop in an OpenAI, Claude, or Gemini node alongside HTTP calls, database writes, Slack messages, and human approvals, all in the same visual editor.
The problem with AI integration
Adding AI to a pipeline usually means another API integration, another deployment, another set of credentials to manage and rotate.
Most AI integrations are stateless. Every call starts fresh with no context about previous interactions, which limits what you can build.
You wire your pipeline to one provider's SDK. Switching from GPT-4.1 to Claude means rewriting integration code, not just changing a config value.
AI responses go straight to production. There's no built-in way to add a human review step before the model's output triggers real actions.
How Tendrl solves it
AI nodes are first-class citizens in the workflow editor, with the same drag-and-drop as every other step.
Pick your model per step
Each AI node in your workflow can use a different provider and model. Use Claude for nuanced classification, GPT-4o for fast generation, and Gemini for cost-effective summarization, all in the same workflow, each with its own temperature and token settings. These are just the common ones: a node accepts any valid model name its provider offers, so a new release works the day it ships.
Use conversation memory
AI nodes can maintain persistent memory across workflow runs. A support-ticket classifier remembers the last 100 interactions. A summarizer builds context over time. Memory is per-node, with configurable TTL (1 minute to 30 days) and automatic trimming when it reaches the limit.
Template prompts with real data
Prompts use Jinja2 templates so you can inject data from any previous workflow step. Pass in the webhook payload, a database query result, or a previous AI response, so the model sees exactly what it needs.
Classify this support ticket into one of: billing, technical, account, other.
Subject: {{ payload.subject }}
Body: {{ payload.body }}
Customer plan: {{ steps.lookup.output_payload.plan }}
Respond with JSON: {"category": "...", "confidence": 0.0-1.0, "summary": "..."}Give AI access to external tools via MCP
Attach MCP servers to AI nodes so the model can call external tools during generation. Your AI node can query a database, look up a customer record, or check inventory, then use that information to generate its response. Multiple MCP servers can be attached to a single node.
Add guardrails and act on the result
After the AI responds, use the same workflow tools you'd use for any other data: Python snippets to parse and validate the output, if/else nodes to route based on confidence scores, and approval nodes to require human review before high-stakes actions.
AI meets the full platform
AI nodes combine with everything else in Tendrl. Here's what makes that powerful.
Strand exposes an MCP endpoint so Claude Desktop, Cursor, and ChatGPT can create workflows, manage devices, and query run history through natural language.
Classify sensor anomalies, generate maintenance recommendations, or predict failures, directly in the same workflow that receives device data from Contact.
Process arrays of items through an AI node with configurable batch sizes. Classify 100 support tickets or summarize 50 documents, with batch delays to stay under rate limits.
AI provider credentials are stored in the encrypted vault. Auto-redacted from logs, API responses, and workflow output. Never exposed in plain text.
Every AI response includes token usage (prompt + completion). Track consumption per workflow, per node, or across your account.
All three providers return the same output structure (response, model, provider, tokens). Switch models without changing downstream nodes.
What people build
Webhook receives a ticket → Claude classifies category and urgency → if/else routes to the right Slack channel → ticket filed in your helpdesk via the HTTP connector with labels and priority set by AI.
Scheduled workflow fetches new documents from S3 → GPT-4o summarizes each one → summaries posted to a Google Sheet → team gets a Slack digest with links.
Form submission triggers workflow → Gemini generates a draft response → human approval node sends draft for review → approved text sent via email to the customer.
IoT data arrives from Contact → Python snippet computes rolling stats → readings outside 2σ sent to Claude for analysis → alert posted with natural-language explanation.
Email webhook delivers invoice PDFs → GPT-4o extracts line items, totals, and vendor info → Python validates the structure → results written to MongoDB.
Gemini Flash for cheap initial classification → Claude for nuanced cases the first model flags → results merged and posted to your CRM through the HTTP connector with confidence scores.
Example: AI support triage with human review
End-to-end from incoming ticket to routed, actioned issue, with AI classification and manager approval for escalations.
Claude Sonnet receives the ticket subject, body, and customer plan. It returns a JSON object with category, urgency, confidence score, and a one-line summary. Conversation memory means it improves with context over time.
If confidence is below 0.8 or urgency is "critical", the workflow pauses for human approval via Slack. The approver sees the AI's classification and summary and can override the category before it continues.
The workflow files the ticket in your helpdesk through the HTTP connector, with the AI-assigned category as a label, the summary as the description, and priority mapped from urgency. A Slack message goes to the relevant team channel with a link to the ticket.
Add AI to your workflows in minutes
Bring your own API key. The free tier includes 10 workflows, 550 runs/month, and 5,000 compute units.