Respell

Converts natural language task descriptions into executable workflow definitions through an LLM-powered intent parser that maps conversational instructions to workflow nodes and connections. The system interprets user intent (e.g., 'send me a Slack message when a new email arrives in Gmail') and translates it into a directed acyclic graph of actions, conditions, and data transformations without requiring users to manually construct the workflow graph.

Abstracts LLM provider APIs (OpenAI, Anthropic, Google, Ollama, etc.) behind a unified interface, allowing workflows to invoke different LLM providers with consistent prompting patterns and parameter mapping. The system handles provider-specific request formatting, token counting, rate limiting, and response parsing, enabling users to swap providers or use multiple providers in a single workflow without modifying workflow logic.

Enables teams to share workflows and collaborate on workflow development through role-based access control that defines permissions for viewing, editing, and executing workflows. The system tracks workflow ownership, manages team access, and provides audit logs of who made changes and when, enabling teams to collaborate safely without requiring shared credentials or manual permission management.

Allows users to embed custom code (JavaScript, Python) within workflows to perform transformations or logic that cannot be expressed through pre-built actions or LLM evaluation. The system executes custom code in a sandboxed runtime environment with access to workflow context (previous step outputs, input parameters) and provides error handling and timeout protection to prevent runaway code from blocking workflow execution.

Provides a library of pre-built workflow templates for common automation scenarios (lead qualification, customer onboarding, support ticket routing, etc.) that users can instantiate and customize. Templates include pre-configured triggers, actions, and logic that users can modify to fit their specific needs, reducing time to deployment and providing reference implementations for best practices.

Maintains stateful conversation context across multiple user interactions, enabling agents to remember prior messages, extract relevant context, and make decisions based on conversation history. The system manages conversation state (message history, extracted entities, decision context) in a structured format, allowing agents to reference prior turns and build coherent multi-step interactions without requiring users to re-provide context.

Defines workflow entry points through declarative trigger configurations that listen for external events (webhook payloads, scheduled times, manual invocations, or provider-specific events like new emails or Slack messages) and automatically instantiate workflow executions when trigger conditions are met. Triggers are configured through a schema-based interface that maps event properties to workflow input parameters without requiring code.

Provides pre-built connectors for popular business tools (Slack, Gmail, Notion, HubSpot, Salesforce, Google Sheets, etc.) that expose tool-specific actions as workflow nodes without requiring users to write API calls. Each connector includes action templates (e.g., 'send Slack message', 'create Notion page', 'update HubSpot contact') with parameter mapping, authentication handling, and response normalization, enabling workflows to interact with external tools through a consistent interface.

Enables workflows to branch execution paths based on conditions evaluated by LLMs or traditional rule engines. The system supports both explicit rule-based conditions (if X equals Y, then branch A) and implicit LLM-evaluated conditions (if the sentiment is positive, then branch A), allowing workflows to make intelligent decisions based on unstructured data without requiring users to write conditional logic code.

Provides visibility into workflow execution status, logs, and error states, with built-in retry mechanisms for failed steps. The system tracks execution progress through each workflow node, captures logs and error messages, and automatically retries failed steps with configurable backoff strategies (exponential backoff, fixed delay, etc.) without requiring manual intervention or workflow re-execution.

Transforms and maps data between workflow steps using schema-based extraction that parses unstructured data (emails, documents, API responses) into structured formats. The system uses LLMs or rule-based parsers to extract relevant fields from unstructured inputs and map them to downstream tool schemas, enabling workflows to normalize data across different tool formats without requiring users to write transformation code.

Maintains version history of workflow definitions, allowing users to view, compare, and rollback to previous versions. The system tracks changes to workflow structure, node configuration, and trigger settings, enabling users to revert to known-good versions if a workflow update introduces errors or unintended behavior without losing execution history or requiring manual workflow reconstruction.

Enables users to test workflows before deploying to production by executing them in a simulation mode that processes real or synthetic data without triggering external side effects. The system captures execution logs, output data, and error states, allowing users to validate workflow logic and identify issues before production deployment without risking unintended actions (sending emails, creating records, etc.).