---

Overview

casen-worker-ai is an official casen CLI plugin that brings AI-powered task processing into Camunda 8 workflows. It subscribes to four job types and uses the Anthropic Claude API to complete each job with structured output.

This is the right tool when you need AI logic in a process but don't want to build a full microservice — start the worker once and any process in your cluster can delegate AI tasks to it.

Why a worker instead of the HTTP connector?

The built-in Camunda HTTP connector can call any API but cannot:

• Validate or parse the response and fail the job if the output is malformed
• Route Anthropic rate-limit errors (429) to Camunda's retry budget with back-off
• Throw a typed BPMN error that an error boundary event can catch

The worker handles all three via failJob and throwJobError, giving your process model clean error paths.

Installation

casen plugin install casen-worker-ai

Set your Anthropic API key before starting:

export ANTHROPIC_API_KEY=sk-ant-...
casen ai-worker

Configuration

Environment variable	Default	Description
ANTHROPIC_API_KEY	required	Your Anthropic API key
AI_MODEL	claude-3-5-haiku-20241022	Model ID to use
AI_MAX_TOKENS	1024	Token budget per call
AI_TIMEOUT_MS	60000	HTTP timeout in ms

Operations

com.bpmnkit.ai.classify — Classify text

Classifies text into one of the provided categories.

Input variables:

Variable	Type	Required	Description
input	string	✓	Text to classify
categories	string[]	✓	Allowed category names
context	string	—	Optional domain context

Output variables: category, confidence (0–1), rationale, aiModel, processedAt

BPMN error codes: AI_INVALID_CATEGORY, AI_PARSE_ERROR, AI_API_ERROR, AI_INVALID_INPUT

<bpmn:serviceTask id="ClassifyTicket" name="Classify ticket">
  <bpmn:extensionElements>
    <zeebe:taskDefinition type="com.bpmnkit.ai.classify" retries="3" />
    <zeebe:taskHeaders>
      <zeebe:header key="retryBackoff" value="PT30S" />
    </zeebe:taskHeaders>
    <zeebe:ioMapping>
      <zeebe:input source="=ticket.body" target="input" />
      <zeebe:input source='=["billing","technical","cancellation","general"]' target="categories" />
    </zeebe:ioMapping>
  </bpmn:extensionElements>
</bpmn:serviceTask>

com.bpmnkit.ai.summarize — Summarize text

Summarizes text to a target length and style.

Input variables:

Variable	Type	Required	Description
input	string	✓	Text to summarize
maxWords	number	—	Target word count (default: 100)
style	"bullet"	"paragraph"	—

Output variables: summary, wordCount, aiModel, processedAt

com.bpmnkit.ai.extract — Extract structured fields

Extracts named fields from unstructured text. Missing fields are reported but the job still completes — use a gateway on =missingFields to decide whether to auto-proceed or route to manual review.

Input variables:

Variable	Type	Required	Description
input	string	✓	Unstructured text
fields	string[]	✓	Field names to extract
schema	object	—	Type hints per field, e.g. { amount: "number" }

Output variables: extracted (object), missingFields (array), aiModel, processedAt

com.bpmnkit.ai.decide — Make a boolean decision

Answers a yes/no question based on context and an optional policy statement. Use confidence at a gateway to route low-confidence decisions to human review.

Input variables:

Variable	Type	Required	Description
question	string	✓	The yes/no question
context	string	✓	Relevant facts
policy	string	—	Natural language policy the model must apply

Output variables: decision (boolean), rationale, confidence (0–1), aiModel, processedAt

Example workflow: Credit decision

Start Event
  → summarize (summarize application notes, maxWords=150)
  → decide     (question="Approve loan?", policy="Score ≥ 680 and amount ≤ 50k")
  → Gateway: decision=true and confidence≥0.8 → Auto-approve
             decision=false and confidence≥0.8 → Auto-deny
             default                           → Human review

Error handling

Add boundary events to your service tasks to handle AI failures gracefully:

• Error boundary catching AI_PARSE_ERROR → route to a manual input fallback task
• Timer boundary (PT5M) → escalation path if the worker is not running
• Set retries="3" and retryBackoff="PT30S" on the task definition — the worker signals Camunda to retry via failJob on rate limits and network timeouts

---

Related Packages

Package	Description
@bpmnkit/core	BPMN/DMN/Form parser, builder, layout engine
@bpmnkit/canvas	Zero-dependency SVG BPMN viewer
@bpmnkit/editor	Full-featured interactive BPMN editor
@bpmnkit/engine	Lightweight BPMN process execution engine
@bpmnkit/feel	FEEL expression language parser & evaluator
@bpmnkit/plugins	22 composable canvas plugins
@bpmnkit/api	Camunda 8 REST API TypeScript client
@bpmnkit/ascii	Render BPMN diagrams as Unicode ASCII art
@bpmnkit/docspack	BPMN Kit docs as an offline docspack package for AI agents
@bpmnkit/ui	Shared design tokens and UI components
@bpmnkit/profiles	Shared auth, profile storage, and client factories for CLI & proxy
@bpmnkit/operate	Monitoring & operations frontend for Camunda clusters
@bpmnkit/connector-gen	Generate connector templates from OpenAPI specs
@bpmnkit/connectors	Camunda 8 OOTB connector catalog and deterministic template application
@bpmnkit/cli	Camunda 8 command-line interface (casen)
@bpmnkit/proxy	Local AI bridge and Camunda API proxy server
@bpmnkit/patterns	Domain process patterns for BPMNKit AIKit
@bpmnkit/reebe-wasm	WebAssembly BPMN engine for browser simulation
@bpmnkit/worker-client	Thin Zeebe REST client for standalone workers
@bpmnkit/cli-sdk	Plugin authoring SDK for the casen CLI
@bpmnkit/create-casen-plugin	Scaffold a new casen CLI plugin in seconds
@bpmnkit/casen-report	HTML reports from Camunda 8 incident and SLA data
@bpmnkit/casen-worker-http	Example HTTP worker plugin — completes jobs with live JSONPlaceholder API data

License

MIT BPMN Kit — made by u11g