AI & agents

Pipory is AI-native throughout the builder, not just in a chat panel: model calls are first-class nodes, a Branch condition can be a plain-English statement instead of an operator, config fields can draft themselves, and a chat assistant explains what you've built.

AI nodes

Three core nodes call a model with a templated system and user prompt, using your own provider credential, and store the generated text downstream as {{yourVariableName.text}}:

• OpenAI
• Anthropic
• Gemini

Six more model providers ship as app-integration nodes (see the AI providers category): Groq, Mistral, Cohere, Perplexity, ElevenLabs, and Replicate. Each needs its own API-key credential the same way any other app node does - see Credentials.

AI Transform (structured extraction)

The AI Transform node turns free text into clean structured JSON: give it an instruction ("extract the sender's name, company, and intent"), point it at an upstream field, and optionally hint at the output shape. It calls Claude with your Anthropic credential and stores the parsed object downstream - the standard way to turn a messy inbound email or form response into {name, company, intent} you can branch and route on.

Natural-language condition

The Branch node has an AI condition mode: instead of an operator and a comparison value, write a plain-English statement - "the email sounds urgent" - and Claude judges it true or false against the current context. Useful for routing on fuzzy signals a strict eq/contains/regex can't express.

AI autofill on config fields

A sparkle button next to certain Handlebars-templated fields (currently HTTP Request, Email, and Slack) opens a one-line prompt - describe what the field should do, and Claude drafts a template referencing your workflow's real upstream variableName fields, resolved straight from the canvas graph.

MCP Tool node

Connects to an external MCP server over streamable HTTP and calls one of its tools, with Handlebars-resolved arguments and the result merged into context - the node lists the server's available tools and invokes the one you pick. This is the inverse of exposing Pipory workflows as MCP tools: it lets a workflow call out to any MCP-compatible tool server.

AI Agent node

A tool-calling agent node: give it a system prompt and a user prompt, and it runs a loop with Claude that can call an HTTP tool and/or a run-context tool (each an opt-in toggle) until it has an answer or hits its iteration cap. Configurable per node (default 5 iterations, hard-capped at 10 regardless of configuration, so a stuck loop can't turn into unbounded API spend), with live streaming status on the canvas like any other node.

AI workflow assistant: understand, build, debug

A chat panel lives in the editor (bottom-right), using your own Anthropic credential, with three capabilities:

• Understand - ask what the open workflow does, why a particular node is configured a certain way, or how the pieces connect. It reads the current workflow's nodes, edges, and configuration (credential ids and secrets are always redacted before anything reaches the model) and answers in plain English.
• Build/edit - ask it to create a workflow from a prompt on an empty canvas, or add a step to an existing one ("when a Stripe payment comes in, post to Slack"). It proposes the change as a set of mutations (add/update/remove a node, add/remove a connection) rendered as a preview on the canvas; nothing changes until you click Apply.
• Debug - ask it about a failing run. It pulls the latest execution and its per-node logs (see Executions) for the open workflow, explains what went wrong, and can offer a Build/edit-style patch when the fix is a configuration change.

Related

• Workflows - how a node's variableName output gets referenced by later nodes.
• Flow control - the Branch node's rules mode, alongside its AI condition mode, and the Approval node.
• Credentials - connecting the API key each AI node needs.
• Public API & MCP - exposing your own workflows as MCP tools.
• Node reference: AI providers