Skip to content

AI Integration Overview

EventConductor is, to our knowledge, the first workflow engine with native Model Context Protocol (MCP) integration.

Every module exposes its domain as MCP tools. Any MCP-compatible AI client — Claude Desktop, a custom chatbot, an internal copilot — can connect and operate the engine in natural language, with no custom integration code.

AI client (Claude Desktop, chatbot, copilot…)
│ MCP over SSE
▼
ia-agent-service (port 8095)
│ opens a fresh SSE connection per prompt
├──► orchestrator MCP server (port 8105) — workflow engine tools
├──► forms-engine MCP server (port 8106) — form definition & execution tools
└──► booking-service MCP server (port 8108) — domain-specific business tools

The process tools have REST twins on the same application services — start, read, list, cancel, pause, resume, retry, restart and retry-step under /workflow/api/processes — so an operation behaves the same whichever door it comes through: the same status rules, the same idempotency on the business key, and (with workflow.security.flow-authorization.enabled) the same definition requirements. The REST API answers synchronously what a tool reports in prose: 404 for an unknown process, 409 for a status that does not allow the command.

ToolDescription
listProcessesProcess instances with status and completion %; optionally filtered by workflowDefinitionId, status, tag and query — search-attribute conditions such as customerId = "C-9" AND amount > 100 (then one page: page, size)
createBatchOperation / listBatchOperations / getBatchOperation / cancelBatchOperationCancel, retry, restart, pause, resume, set variables or retry from a step for every process a filter matches, in the background; follow its progress, skip reasons and errors; cancel it (guide)
migrateProcessesInBatchA MIGRATE batch: move every live process of a definition (of one version, fromVersion) onto another version with one step mapping
getProcessDetailsFull process detail: variables + all step executions
findProcessByBusinessKeyLook up a process by its business key
getProcessLogsAudit trail and log messages for a process
listInFlightProcessesProcesses in flight (PENDING, RUNNING, PAUSED) from the process-index read model — one indexed read; empty when workflow.projection.enabled is off
countProcessesByStatusProcess counts by status across the fleet, from the process-index read model; empty when the projection is off
retryProcessPick a stopped process up where it stopped: the failed (or cancelled) steps run again, the ones that succeeded are left alone. ERROR or CANCELLED only
restartProcessRun a stopped process again from the beginning, the steps that already succeeded included, with the variables it was created with. ERROR or CANCELLED only
setProcessVariablesChange a live process’s variables (set, remove) — an audited admin operation; reason mandatory (guide)
revealProcessVariableThe value of one sensitive variable (shown as •••• by every other tool) — needs a workflow.security.sensitive.reveal-roles role and a reason; audited (guide)
completeStepManuallyComplete a step execution by hand with output variables, as if its worker had answered; the flow carries on
retryProcessFromStepRun a live process again from one of its steps: it and everything downstream start over
dryRunProcessMigration / migrateProcessPlan, then carry out, moving a live process onto another version of its definition — a step mapping, variables, the reason (guide)
listIncidents / getIncidentProcesses / retryIncidentFailures grouped by definition, step and error family; their processes; retry a whole group — as a RETRY batch operation, whose id it returns
pauseProcess / resumeProcessPause a PENDING or RUNNING process (in-flight work finishes, successors held, clocks frozen) and resume it
pauseWorkflow / resumeWorkflowPause/resume a whole definition: all its processes, with new instances born paused
sendMessageSend an external message to resume processes waiting on a WAIT_FOR_MESSAGE step
updateProcess / updateWithStart / listProcessUpdatesApply a declared update to a running process (by id or business key) and get its result or why not; the same, starting the process when the key has none; a process’s updates (guide)
getWorkflowAnalyticsPer-definition analytics: counts by status, rates, throughput per day, avg/p95 durations, bottleneck step (guide)
findBottleneckWhere processes get stuck: slowest step, steps with waiting instances, steps with failures
importWorkflowDefinitionsFromGitPull and upsert workflow JSON files from Git
listRunsOfBusinessKeyEvery run that has had a business key, oldest first — a CONTINUE_AS_NEW chain (linked by continuedFrom/continuedAs) or the runs a businessKeyPolicy let reuse it
listWorkflowDefinitionsThe definitions (id, name, status, whether they accept new instances) — what a schedule can start
listSchedules / getScheduleEvery schedule, optionally of one workflow: cron (and in words), zone, variables, source, whether it is editable, paused, next starts
createSchedule / updateSchedule / deleteScheduleManage the schedules created through the engine (source UI/API). A schedule declared in a definition or a cron.yaml is owned by its file and refused here
pauseSchedule / resumeSchedulePause or resume any schedule; the pause is kept by id and survives re-imports
runScheduleNowStart the schedule’s workflow once, now, with its variables
createEventTrigger / updateEventTrigger / deleteEventTrigger / exportEventTriggerYamlCreate a trigger (source UI/API) that starts a workflow on a domain event, edit or delete one created that way, and get its events.yaml entry to promote it to git (guide)
listEventTriggersEvery event trigger, optionally of one workflow — declared in the definition, in an events.yaml or created through the engine: destination and topic, type, filter, business key, variables, source, status (guide)
pauseEventTrigger / resumeEventTriggerPause or resume a trigger; while paused, matching events start nothing and are not replayed
listReceivedEventsThe latest received domain events and what each did: started a process, delivered a message, or the error
publishInboundEventInject a domain event on a destination (for tests and systems without Kafka); returns what it started and delivered
exportScheduleYamlThe schedule’s cron.yaml entry, id included, to promote it to git (lifecycle)
ToolDescription
listFormsAll form definitions with field count
listFormExecutionsAll pending/completed user tasks
getFormExecutionFull detail of a form execution including submitted values
importFormsFromGitPull and upsert form JSON files from Git
ToolDescription
listRulesAll rule definitions in the catalog
getRuleFull definition of a rule as canonical JSON
saveRuleCreate or update a rule definition (JSON or YAML, schema-validated)
validateRuleValidate a rule definition without saving it
deleteRuleDelete a rule definition by id
evaluateRuleEvaluate a rule against a JSON object of facts
importRulesFromGitPull and upsert rule JSON/YAML files from Git

Custom domain tools (example: booking-service, port 8108)

Section titled “Custom domain tools (example: booking-service, port 8108)”
ToolDescription
createBookingCreate a new booking for a lead
listBookingsAll bookings with status
getBookingFull booking detail
changeBookingStatusChange booking status (Pending / Confirmed / Cancelled)
Natural languageWhat happens
“What is the status of order 123?”Queries process by business key, returns status + variables
“Retry all failed processes from today”Calls retryProcess for each ERROR process
“Show me the pending user tasks for the onboarding workflow”Lists form executions filtered by workflow
“Import the new workflow definitions from Git”Triggers importWorkflowDefinitionsFromGit
“Cancel booking B-456 and tell me why it failed”Changes booking status + reads process logs

The agent is designed to be resilient to MCP server failures: if one MCP server is down, the others continue working normally. Each server self-describes its domain via a system-context MCP Prompt, so the LLM always has up-to-date context without any code changes.

DecisionReason
Fresh SSE connection per promptA persistent SSE connection breaks if the MCP server restarts. Opening a fresh connection per request makes the agent resilient to server restarts.
Dynamic system promptEach MCP server exposes a system-context MCP Prompt. The agent reads it on each request and combines it with a base prompt — no code changes needed when servers change.
Anthropic Prompt CachingThe system prompt is cached server-side (Anthropic cache). Cached tokens cost ~10% of normal price, significantly reducing cost when the system prompt is stable.
Conversation historyThe last 5 exchanges are stored per session (Caffeine cache), giving the LLM context across a conversation.