Camunda

Agents act through the BPMN activities a builder places in an ad hoc sub process, so any service task, connector call or sub process becomes a tool the model can choose. Camunda offers more than 400 prebuilt connectors, including Salesforce, SAP OData and RFC, HubSpot, Slack, Microsoft Teams, GitHub, Twilio, WhatsApp, Google Drive and Sheets, Amazon S3, DynamoDB and Lambda, plus generic REST, SOAP, GraphQL and SQL. Credentials sit in secrets, a connector SDK builds custom connectors, and MCP client connectors let an agent discover and call tools on outside MCP servers.

Sourcedocs.camunda.io/docs/components/connectors/out-of-the-box-connectors/available-connectors-overviewread 2026-10-12

Processes run on the Zeebe engine as BPMN, with gateways, timers, events, sub processes and DMN decisions. An AI agent is an addressable loop inside that process. The model picks the next activity from an ad hoc sub process, the engine runs it as a governed step, and the result goes back to the model until the goal is met. Several agents can share one process, agents built in LangGraph, Amazon Bedrock or custom code can be connected and tracked as agent instances, and predictable steps run straight through while agents handle the judgment calls.

Sourcedocs.camunda.io/docs/components/agentic-orchestration/ai-agentsread 2026-10-12

Agents can draw on company knowledge kept in a vector store. The Vector Database connector splits documents into chunks, embeds them with Amazon Bedrock Titan, Azure OpenAI, Google Vertex AI or OpenAI models, and writes them to Amazon OpenSearch, OpenSearch, Elasticsearch, Azure AI Search or Azure Cosmos DB. At run time it retrieves the most relevant passages by semantic similarity, so an agent can look up policies and procedures, search a large product catalog, reuse answers to past support questions or quote the exact policy behind a decision. The store grows as new documents are embedded.

Sourcedocs.camunda.io/docs/components/agentic-orchestration/rag-ai-agentsread 2026-10-12

People sit in the process as user tasks, and Tasklist shows each assigned task whether the step before it ran automatically or an AI agent produced the result a person now acts on. A user task can be one of an agent's tools, so the agent asks for help or sign off inside its loop. For MCP tools, a gateway in front of the client lets listing and read operations through and routes selected calls to a user task for confirmation before they run. Task work can also reach people in Microsoft Teams, and teams can build their own task apps on the API.

Sourcedocs.camunda.io/docs/components/connectors/out-of-the-box-connectors/agentic-ai-mcp-client-human-in-the-loopread 2026-10-12

Camunda runs two identity layers, the Orchestration Cluster admin for runtime work and the management identity for Hub and Optimize, and both connect to an outside OpenID Connect provider such as Entra ID for single sign on, with LDAP and Active Directory on the licensed editions. Authorizations, roles, groups, mapping rules and tenants decide who can start, see and change processes, tasks and other resources. The company holds SOC 2 Type II, ISO 27001 and TISAX and is listed with the Cloud Security Alliance on its trust center.

Sourcetrust.camunda.comread 2026-10-12

Operate shows each agent instance's state, such as thinking or calling a tool, its usage metrics and its conversation history grouped by loop iteration, including the tools it picked and the results it got back, for native and connected external agents alike. Health indicators and limit warnings flag an agent that keeps calling the same tool or nears its configured limits. Every process and decision instance can be inspected step by step, audit log APIs record user and user task actions, and Optimize builds reports, dashboards and KPI alerts.

Sourcedocs.camunda.io/docs/components/agentic-orchestration/evaluate-agents/monitor-ai-agentsread 2026-10-12

Each agent keeps a short term conversation inside its process instance, passing the last 20 messages to the model by default while always keeping the system prompt. Those messages can live in a process variable, in Camunda document storage with a time to live set to the process's expected lifetime, in Amazon Bedrock AgentCore Memory, or in a custom store on self managed setups. Recall from one process instance to the next relies on an outside service such as AgentCore Memory, which AWS runs.

Sourcedocs.camunda.io/docs/components/connectors/out-of-the-box-connectors/agentic-ai-aiagent-subprocessread 2026-10-12

SaaS customers choose the region for each cluster. On Google Cloud the choices include Belgium, London, Iowa, South Carolina, Montréal, Toronto, Singapore and Sydney, and on AWS they include Frankfurt, Paris, Ohio, Oregon, Virginia and Singapore, each paired with a secondary backup region. Enterprise customers can ask for custom regions. Self managed Camunda runs the whole stack on the customer's own infrastructure, and agent model calls can route through a private endpoint inside the customer's network.

Sourcedocs.camunda.io/docs/components/saas/regionsread 2026-10-12

Camunda Marketplace carries model blueprints and connectors from Camunda, partners and the community, installed straight into a Hub project or downloaded for the desktop modeler. AI agent blueprints such as AI Agent Chat Quick Start and AI Agent Chat With Tools run out of the box on the Camunda provided LLM in SaaS. Partner and community connectors sit outside Camunda's commercial support, and administrators of a self managed Hub can switch the Marketplace off.

Sourcedocs.camunda.io/docs/components/hub/workspace/modeler/modeling/camunda-marketplaceread 2026-10-12

Processes start or move on without anyone pressing a button. Inbound connectors listen on an HTTP webhook, Kafka, RabbitMQ, Amazon SQS and SNS, Amazon EventBridge, Slack, email, Microsoft 365 mail, GitHub and Twilio, and BPMN timer and message events start or resume instances on a schedule or when a message arrives. An MCP start event registers a deployed process as a tool that outside AI clients can call, and each call starts a new instance.

Sourcedocs.camunda.io/docs/components/connectors/out-of-the-box-connectors/available-connectors-overviewread 2026-10-12

The builder picks the provider and the backend for each agent. Anthropic Claude models run directly, through AWS Bedrock or through Microsoft Foundry, OpenAI models run directly, through Foundry or through any OpenAI compatible endpoint, and Google Gemini runs directly or through Google's Enterprise Agent Platform, with custom endpoints for private network setups and a custom provider on self managed installs. SaaS adds a Camunda provided LLM with a free budget for testing, so teams can switch between models such as Claude Sonnet 4.6 before moving to their own account.

Sourcedocs.camunda.io/docs/components/connectors/out-of-the-box-connectors/agentic-ai-aiagent-model-providersread 2026-10-12

The Orchestration Cluster REST API drives processes, instances, user tasks, decisions, identities and audit logs, with SDKs for Java, TypeScript, Python, C# and PHP and the c8ctl command line tool. The Orchestration Cluster MCP Server opens operational work such as incidents, user tasks and process instances to AI clients under the same authentication and authorization model as the REST API. Older Operate and Tasklist REST APIs remain available for existing integrations.

Sourcedocs.camunda.io/docs/apis-tools/orchestration-cluster-api-mcp/orchestration-cluster-api-mcp-overviewread 2026-10-12

Camunda Process Test runs an agent process as an integration test, keeping the real model in the loop while mocking outside tool calls. Conditional behavior reacts to whichever tools the agent activates instead of one fixed order, and judge assertions send an output and a plain language expectation to a model that rates the match against a set threshold, alongside semantic similarity checks. Tests run through a Spring Boot starter inside the project's own build.

Sourcedocs.camunda.io/docs/components/agentic-orchestration/evaluate-agents/test-ai-agentsread 2026-10-12

Camunda RPA runs Robot Framework scripts on a standalone RPA worker to operate legacy mainframe applications, desktop applications such as ERP systems and websites that expose no API. Each script runs as a step in a process, so screen work on an old system sits in the same flow as API calls, agent steps and user tasks, and scripts are built and tested before they move to production workers.

Sourcedocs.camunda.io/docs/components/rpa/overviewread 2026-10-12