botpress

Botpress abstracts multiple LLM providers (OpenAI, Anthropic, Ollama, etc.) through a unified SDK layer (@botpress/llmz package) that normalizes provider-specific APIs into a common interface. This enables swapping LLM backends without changing bot logic, using a provider registry pattern that maps configuration to concrete implementations. The abstraction handles token counting, streaming, function calling, and error handling across heterogeneous providers.

Botpress provides an IntegrationDefinition class that allows developers to declare integrations (messaging platforms, CRMs, APIs) using a schema-based approach where configuration, actions, events, and channels are defined as TypeScript classes. The framework generates type-safe bindings and automatically handles serialization, validation, and runtime dispatch. Integrations are discovered and loaded via a plugin system that supports 50+ pre-built integrations (Slack, Discord, Telegram, Salesforce, etc.).

Botpress bots maintain conversation state across multiple message exchanges using a context object that persists user metadata, conversation history, and custom variables. The context is passed through the event handler chain, allowing middleware and handlers to read and modify state. State can be stored in memory (for development) or external stores (Redis, PostgreSQL) for production. The SDK provides utilities for serializing/deserializing context and managing conversation lifecycle (start, end, timeout).

Botpress integrates function calling (tool use) by allowing bots to invoke integration actions through LLM-generated function calls. The SDK converts integration action definitions into JSON schemas that are passed to LLMs, enabling models to decide when and how to call actions. The framework handles schema validation, function dispatch, and result formatting. This enables agentic workflows where bots autonomously decide which integrations to invoke based on user intent.

Botpress provides channel-specific message rendering that adapts bot responses to platform capabilities. Bots define messages using a unified format (text, cards, buttons, etc.), and the SDK renders them appropriately for each channel (Slack formatting, Discord embeds, Telegram inline keyboards, etc.). The framework handles platform-specific limitations (character limits, supported media types) and provides fallbacks for unsupported features.

Botpress implements a PluginDefinition class that enables extensible functionality through plugins, with a specialized HITL plugin that orchestrates human handoff workflows. Plugins hook into the bot lifecycle (message processing, event handling) and can intercept, modify, or escalate conversations to human agents. The HITL plugin provides conversation routing, agent assignment, and conversation history management through a standardized interface.

Botpress CLI (@botpress/cli) provides commands to scaffold new bots, integrations, and plugins from templates (empty-bot, hello-world, webhook-message, etc.). The CLI generates boilerplate TypeScript code with proper SDK imports, configuration, and build setup. It handles project initialization, dependency management via pnpm, and provides commands for local development (build, serve) and deployment to Botpress Cloud.

Botpress SDK provides a BotImplementation class that allows developers to define bot logic as event handlers and lifecycle hooks (onMessage, onEvent, onInstall, etc.). Bots are implemented as HTTP servers (via botHandler) that receive events from integrations and dispatch them to handler functions. The architecture supports middleware-style composition where multiple handlers can process the same event sequentially.

Botpress is organized as a pnpm monorepo with ~12 core packages, 50+ integrations, plugins, and example bots in a single repository. The monorepo structure enables shared dependencies, consistent versioning, and cross-package type checking. Developers can modify core SDK, integrations, and plugins simultaneously and test changes locally before publishing. The build system uses TypeScript compilation and bundling to generate distribution artifacts.

Botpress integrations define actions (bot-initiated operations) and events (platform-initiated notifications) using TypeScript classes with schema validation. Actions are type-safe functions that integrations expose to bots; events are typed payloads that integrations emit when external platforms trigger them. The SDK generates runtime validators from TypeScript types, ensuring data integrity at integration boundaries. This enables IDE autocomplete and compile-time type checking for integration interactions.

Botpress includes a knowledge synchronizer plugin that enables bots to retrieve relevant context from external knowledge bases (documents, FAQs, wikis) using semantic search. The plugin integrates with vector databases and embedding models to implement Retrieval-Augmented Generation (RAG), allowing bots to ground responses in factual information. Retrieved context is injected into LLM prompts, reducing hallucinations and improving answer accuracy.

Botpress CLI includes a deploy command that packages bots and integrations and deploys them to Botpress Cloud, a managed hosting platform. The platform handles automatic scaling based on message volume, provides built-in monitoring and logging, and manages SSL certificates and domain configuration. Deployments are versioned, enabling rollback to previous versions. The platform integrates with Botpress integrations, automatically provisioning necessary credentials and webhooks.

Botpress integrations receive messages from external platforms (Slack, Discord, Telegram, etc.) via webhooks. The SDK provides a webhook template that handles incoming HTTP POST requests, validates signatures, and dispatches events to bot handlers. Integrations define channel configurations that specify webhook URLs and authentication methods. The architecture supports both synchronous responses (immediate replies) and asynchronous processing (background jobs).