WisebotAI
Workflows

Workflows

Build visual automations that chain inputs, AI, knowledge base, agent tools, MCP integrations, HTTP, and code—with manual run, webhooks, and schedules.

Workflows

Workflows let you design repeatable automations as a simple diagram: each box is a step, and lines define the order steps run. You can run a flow from the dashboard, trigger it with a webhook, or run it on a schedule.

They complement chat agents: the same knowledge base, built-in tools, and MCP integrations can be used inside a workflow, but the flow runs as a scripted pipeline (with a clear result and run history) instead of a back-and-forth conversation.

Why use workflows

  • Consistency — The same steps run the same way every time (manual, scheduled, or event-driven).
  • Composition — Combine inputs, LLM steps, KB retrieval, agent tools, MCP tools, HTTP calls, and sandbox code in one place.
  • Automation — Webhooks connect external systems; schedules run flows at a fixed interval.
  • Visibility — Recent runs show status, errors, and technical detail for support or debugging.

How to build a flow

  1. Add steps — Use the + menu on the canvas to place steps (grouped by purpose: inputs & outputs, apps & code, logic, AI & knowledge).
  2. Connect steps — Drag from the handle on the right of one step to the handle on the left of the next. The graph must stay a valid order (no cycles).
  3. Configure — Click a step to edit its settings in the right panel (display name, keys, prompts, URLs, and so on).
  4. Save — Persist your diagram before running or enabling automations.
  5. Run — Use Run now; manual values for Input steps live on the Run & automate tab when present.

Triggers

TriggerWhat it does
Manual (Run now)You start the run from the dashboard and optionally fill fields that feed Input steps.
Webhook (POST)Another system sends an HTTP POST to your workflow URL with a secret header. JSON body fields populate the input bag when keys match Input steps.
Dashboard APIPOST /api/v1/flow/{workflowId} with org JWT or wb_live_ API key and body { "inputs": { … } }. See Triggering workflows.
ScheduleWhile Scheduled runs is on, the deployment runs the workflow on a timer (about once per chosen interval). The next approximate run time is shown when scheduling is enabled.

Step types (overview)

Data in and out

StepPurpose
InputNamed fields: text, JSON, number, or file upload (stored securely; later steps receive a file reference).
OutputDefines the workflow result (what shows up as the primary output for a completed run).

AI and knowledge

StepPurpose
LLMCalls the configured model with your prompt; optional file text extraction, vision on an image, or slide-outline style Markdown.
KB preflightRetrieves scoped knowledge snippets the same way as before an agent turn (uses your default agent or a chosen agent with linked knowledge files).
KB searchRuns the same vector knowledge search as chat agents. Requires a real conversation thread id and query supplied via your inputs (for flows tied to live chats).

Agents and tools

StepPurpose
Agent turnRuns one full channel agent turn (tools, KB preflight, message save) for a real thread, contact session, and user message—the same stack as WhatsApp/Telegram-style agents.
Agent toolRuns a single built-in or custom tool (search, booking, escalate, resolve, MCP-backed tools, or custom:…) for a thread and session—same execution path as chat tools.
MCP toolCalls a tool on a connected MCP integration (Notion, GitHub, Slack, Google, custom MCP server, and so on). Pick the integration, tool name, and JSON arguments (from an input field or default JSON). Connect integrations under Integrations first.

Integrations and code

StepPurpose
HTTPSGET/POST to an https:// URL. You can substitute Input values into the URL using {{fieldName}} placeholders.
CodeRuns bash, Python, or Node in an E2B sandbox when your deployment has E2B_API_KEY configured. Receives context (inputs, trigger, payload, input bag).
Page crawlHeadless browser fetch of an https:// URL — returns title and readable markdown (good for JS-heavy pages). Requires E2B.
Computer useLaunches apps in an E2B desktop sandbox (browser, IDE), optional bash script, optional screenshot. Requires E2B.
TransformReads one field from upstream data using a dot path (for example bodyPreview or nested.field).
ConditionBranches the graph based on input values or expressions.
LogEmits a log line (in the deployment logs or sandbox, depending on configuration).

Run history and debugging

  • Each Run now, webhook, or scheduled execution appears in recent runs with status, timestamps, and error detail when something fails.
  • Webhook and schedule toggles live under Run & automate (generate or rotate the secret there; keep it out of client-side code).

Requirements and limits (operators)

  • LLM steps need a model provider configured for your organization (the same kind of setup that powers agents).
  • Code and some outbound HTTPS paths may require sandbox credentials to be enabled for your workspace—your administrator configures this.
  • KB search, Agent turn, and Agent tool steps that rely on a conversation must receive a thread (and usually a contact session) that belongs to your organization—they are not a substitute for creating chats out of thin air.
  • MCP tool requires a connected integration that supports remote tool calls. Older connector types that do not expose tools the same way are not supported (errors match what you see when using MCP elsewhere).

Analytics

Metrics for conversations, channels, and agent performance.

Custom tools + MCP validation pack

Build and validate four high-impact custom tools, then verify custom MCP and REST tool calling end-to-end.