Botpress AI Agents: Studio, Botpress Viber and the ADK Explained

A Botpress AI agent is a bot that follows a prompt: you write instructions, and a large language model decides what to say and which tools to call. Botpress gives you three ways to build one, on the same plans and the same price list: the Autonomous Node in Botpress Studio, the plain-language Botpress Viber, and the TypeScript ADK. The Free plan covers 3 AI agents and 100 conversations, while Plus, at $150 a month billed annually or $189 a month billed monthly, lifts the agent count to unlimited ^[1]. Workspaces created before mid-May 2026 keep their earlier plan for now, so older price pages and tutorials may not match yours. Pick the builder by who will maintain the agent.

What does the thinking inside a Botpress agent

LLMz, Botpress's own inference engine, sits under the Studio Autonomous Node and the ADK's execute() loop. Instead of firing tool calls one at a time, it turns plain-language instructions into TypeScript code that runs in a single request. In Studio, its version can be changed in Bot Settings, where Botpress recommends the latest stable release. Tools therefore look like functions: Studio Actions are reusable TypeScript functions only Autonomous Nodes can call.

Every new Studio bot already contains an agent. The Main Workflow holds a Start Node, an Autonomous Node and an End Node, and the Autonomous Node arrives with built-in tools to think, search knowledge, set a reminder and send a message, plus browser.webSearch for the web and hitl.startHitl for a human handoff. Its Search Knowledge card queries every knowledge base by default. What it cannot do by default is read or change your variables: read and write access is granted per variable ^[2].

Studio: let the Autonomous Node talk, let workflows decide

Studio mixes two node types freely. Standard Nodes run cards in sequence and use no AI by default; the Autonomous Node lets the model choose what to say and which tools to run, looping instead of having an outgoing transition. The pattern Botpress teaches is simple. Keep judgment calls in the Autonomous Node and move anything that must not improvise into workflows, which bypass the model's logic entirely, as Botpress advises for financial transactions and escalation to a human.

The handover from agent to workflow works two ways: a keyword transition condition, or an instruction such as sending anyone who asks for a person to the handoff workflow. Inside workflows, the AI Transition card sorts user text into categories without training phrases. An Autonomous Node can also run non-conversationally, executing logic without replying. Each node can override the bot's default model, and Botpress's docs advise a high-performing one, because smaller models may truncate the tool definitions the agent relies on.

Botpress Viber: an agent described in plain language

Botpress Viber drops nodes, cards, visual workflows and code. You describe the agent to Vibe, the builder agent, and it turns the request into configuration. Behavior comes in three layers. Instructions cover identity, communication, memory and safety and always apply; Playbooks hold reusable procedures; Escalation defines the handoff to a human, which stays off until you enable it and then lands in Botpress Desk. An optional pre-handoff procedure can collect context first.

Knowledge comes from a website, FAQs, files and Web Search, which starts switched on for the whole web, so narrow it if answers should come only from your material. Tools come from catalog integrations, custom HTTP Request tools (no authentication, bearer token, basic authentication or API key) and MCP server tools. Connecting an integration is not enough. The agent calls an action only after Instructions or a Playbook reference it.

Testing takes two forms. Question tests rate a single answer with its cited sources, while flow tests run a multi-turn scenario against one Playbook with the tools mocked; Monitor then shows Conversations, Simulations and Gaps. One limit matters while you evaluate: Botpress Viber workspaces on the Free plan get a daily build allowance, and once it is used up, an in-product paywall appears.

The ADK: when the agent lives in your repository

The ADK is a command-line tool plus a TypeScript framework on Node.js 22 or later. adk init scaffolds a project, adk dev runs it locally against your own dev bot, and adk deploy ships it to the production bot on Botpress Cloud. The core is execute(), a loop that ends when the model messages the user, triggers an exit or reaches the iteration limit, which defaults to 10 and accepts 1 to 100 ^[3]. Chat mode is the default; worker mode runs in the background without posting to the conversation.

Each call can name its own model with an ordered list of fallbacks. Workflows run independently of conversations and can be scheduled with cron, started in code or called by the model as a tool. They time out after 5 minutes by default. Longer jobs split into steps, persisted checkpoints that retry automatically, by default for up to 5 attempts, before the error fails the workflow (unless the code catches it). Projects export as a portable .adk archive, which is how a code-first team moves an agent between directories.

The quality tooling is where the ADK pulls ahead. Evals simulate conversations and assert on replies, tool calls, state and timing, including checks scored by an LLM judge, and traces export to an OpenTelemetry collector. Agent(0), the assistant in the ADK's console, edits actions, runs CLI commands and debugs, while adk init installs ADK skills for Claude Code, Cursor and Codex. The Agent Map view is experimental.

Which model runs your agent, and what it spends

Studio supports models from OpenAI, Anthropic, Google, xAI, GLM and Mistral (through an integration), plus open models through Fireworks AI, Cerebras and Groq. As of October 2026, the default OpenAI model is GPT-6.1 Sol and the default Anthropic model is Claude Sonnet 5.5; new Studio bots use GPT-6 Luna for the Best setting and Gemini 3.1 Flash Lite as the fallback . Bringing your own model through the SDK is Enterprise only, arranged with sales.

Model calls draw on AI usage bundled into each plan, charged at the provider's cost with no markup. Plus includes $25 of AI usage alongside its 250 conversations a month, and Team includes $150 alongside 1,500; by our arithmetic both work out to $0.10 per included conversation. At 95% of the AI spend limit on paid plans, Botpress buys a $10 credit grant at cost, and grants last up to 12 months but are not refundable. Emulator tests count toward the conversation quota and draw on AI usage, as our Studio build guide explains.

To see where the money goes, Analytics shows which models are fastest, return errors, or consume the most tokens and AI spend. For per-turn numbers, the After Turn End hook reports tokens, billed tokens and AI spend in nanodollars (divide by 1,000,000,000 for dollars). Botpress caches AI responses and says this can cut query cost by about 30%. Knowledge bases are usually the largest cost driver, which is why our knowledge base guide covers narrowing searches.

Guardrails for an agent that takes actions

Botpress says its agents can issue refunds, correct orders and update accounts under the same conditions a human agent would, with configurable approval thresholds that send ambiguous or high-value cases to a person first. Read that as a design you configure. It is not a default. The Policy Agent enforces rules such as banned words or not discussing competitors, RAG Safety shows which knowledge chunks were selected, and encrypted configuration variables keep API tokens out of an Autonomous Node's reach.

Escalation is the other half of safety. In Studio, an instruction or keyword sends the user to a handoff workflow, and the Summary Agent prepares a summary and transcript for the person taking over. In Botpress Viber, enabled escalation goes to Botpress Desk; in the ADK, the desk-hitl plugin creates a Desk ticket. The triggers Botpress describes are topic, sentiment, customer tier, an explicit request and a confidence threshold. Set one before launch.

Studio, Botpress Viber or the ADK for your first agent

Choose by ownership. If a support or operations lead will maintain the agent and the job fits into instructions and playbooks, Botpress Viber keeps everything in plain language, with tests built in. If some steps must never improvise, Studio lets the Autonomous Node converse while workflows hold the rules. If developers want cron workflows, evals and traces, the ADK fits. All three share one price list, so switching builders needs no new plan; teams weighing a rival that also builds from playbooks and workflows can read Botpress vs Voiceflow.