Workflows API Overview - Enginy API Docs

The Workflows API lets you build, publish, and run Advanced Workflows — the graph automations that import contacts and companies, enrich them, branch on their data, and act on the result — entirely over REST (and, through the same operations, over MCP). This page is the mental model. If you just want to ship something, jump to the Cookbook; for the full request/response contract of each call, see the interactive Workflows endpoints in the Endpoints section.

This is the API view of workflows. For the visual editor, templates, and per-step product reference, see the Workflows product guide. Both drive the same underlying workflow — a plan you publish over the API opens and runs identically in the app.

What you work with

Concept	What it is
Plan	The workflow graph you submit: `{ name?, nodes: [] }`.
Block	A single workflow step (import, enrich, condition helper, action). Every `BLOCK` node references a `blockId` from the block catalog.
Node	One vertex in the plan graph. Three kinds: `BLOCK`, `CONDITIONAL`, `KING`.
Draft	The editable working copy of a workflow. Every create/update writes the draft.
Published version	An immutable snapshot of the draft that a run executes. Publishing bumps the version number.
Run	One execution of a published version over a set of records.

Configure an import source from an audience

You can build a recurring search from natural-language criteria such as “100 software engineers in Barcelona city” without supplying a LinkedIn search URL.

Call POST /v1/ai-finder/previews (preview_an_ai_finder_search) with the audience and provider, or reuse the latest matching preview.
Call GET /v1/workflows/import-inputs?previewId=<previewId>&maxCount=100 (get_workflow_import_inputs, scope WORKFLOWS_READ).
Add an import node with the returned blockId and inputs: {}. Review any warnings.
Send importSources: [{ nodeId, ...source }] alongside plan to both validate and create, using the resolver’s short source: { previewId, maxCount } reference. The server applies the exact source inputs before validation and saving, so generated URLs and filter objects do not need to be copied by the model. Validation returns the normalized plan for inspection when references are present.

For example, if the source node has id: "prospects", both request bodies include:

{
  "importSources": [{ "nodeId": "prospects", "previewId": "<previewId>", "maxCount": 100 }]
}

Draft updates accept the same references with a replacement plan, edit operations, or an existing node id. Each reference must target a unique block node whose blockId matches the resolved source. References are consumed during authoring; saved workflows keep the resolved inputs and do not depend on a preview at execution time. Source values replace any conflicting values on that node. Tenant, expiry, and identity access checks apply on every resolution. The resolver generates LinkedIn people/company search URLs and prepares Google Maps, fundraising data, StoreLeads, TheirStack, and supported CRM import inputs. It does not import records, create a destination list, or run the workflow. A scheduled import repeats the search on each run; selecting the previous results list reuses those records instead. Previews expire after 24 hours. If unavailable, regenerate a preview from the original criteria. Unsupported sources return an error rather than guessed inputs. Some CRM import blocks cannot enforce a per-run maximum; their responses include a warning so callers can decide how to proceed. LinkedIn post/event/job URLs and Pipedrive/Odoo source configurations are outside this resolver’s supported search types. For weekly schedules, recurrence.byWeekday uses 0 for Sunday, 1 for Monday, through 6 for Saturday. Include an explicit UTC offset or Z in startsAt so its instant corresponds to the requested wall-clock time in the schedule’s IANA timezone; an offset-free timestamp depends on the server timezone.

The graph model

A plan is a directed acyclic graph. Every node has a stable string id and a next array of the node ids it points to. The empty string "" marks an unwired slot (a dangling branch a draft can tolerate but a publish cannot). The entry node is the single node that no other node points to. It must be a selector or import block — the step that brings records into the workflow.

Node kinds

BLOCK — one workflow step

{ "id": "verify", "kind": "BLOCK", "blockId": "VERIFY_LEAD_EMAIL", "inputs": {}, "next": ["campaign"] }

blockId comes from GET /v1/workflows/blocks.
inputs are the block’s configured inputs — discover them with GET /v1/workflows/blocks/{blockId}.
A BLOCK has exactly one next slot, except WAIT_FOR_APPROVAL, which has two: [approvedTargetId, rejectedTargetId].

CONDITIONAL — branch on record data

{
  "id": "gate",
  "kind": "CONDITIONAL",
  "conditions": [{ "field": "professionalEmail", "operator": "IS_NOT_EMPTY" }],
  "fallbackName": "No email",
  "next": ["campaign", "review"]
}

next has exactly conditions.length + 1 entries: one target per condition (in order), then the fallback target last.
Discover fields and operators with GET /v1/workflows/condition-fields. The catalog includes entity fields plus engagement, LinkedIn, CRM, platform, activity, and run-context fields. Each field reports its compatible operatorIds and fixed valueOptions when applicable.
At most 20 conditions on a single node.

For example, route records that belong to list 55901 with the built-in platform.listIds field:

{
  "field": "platform.listIds",
  "operator": "CONTAINS",
  "value": 55901
}

A condition can also be a random split (deterministic A/B split) instead of a field predicate:

{
  "id": "ab",
  "kind": "CONDITIONAL",
  "conditions": [{ "type": "RANDOM_SPLIT", "percentage": 50 }],
  "fallbackName": "Variant B",
  "next": ["variant-a", "variant-b"]
}

percentage is an integer 1–100; sibling split percentages on one node must sum to ≤ 100.
Any remainder falls through to the fallback branch (so one RANDOM_SPLIT of 50 is a 50/50 A/B split).
Each record gets a stable roll per run — re-running a step never reshuffles records between branches.

Prioritize by group (KING)

{
  "id": "prioritize",
  "kind": "KING",
  "groupBy": "companyId",
  "orderBy": 1234,
  "limit": 10,
  "next": ["campaign", "review"]
}

groupBy chooses the field that forms the groups, orderBy chooses the formula field used to sort each group, and limit is the exact number of records prioritized from every group (default 1).
Lower formula values are prioritized first; records without a value come last. Ties do not expand the limit.
next is [prioritizedTargetId, remainingTargetId].
The example prioritizes up to 10 contacts per company for campaign and sends the remaining contacts to review.
KING remains the stable plan kind even though the user-facing step is named Prioritize by group.

Wiring rules

The next slot count is fixed per node kind: BLOCK → 1 (or 2 for WAIT_FOR_APPROVAL), CONDITIONAL → conditions.length + 1, KING → 2.
Every id in a next array must reference an existing node id, or be "" for an unwired slot.
No cycles. A plan may hold at most 60 nodes.
The entry node is inferred (the node nothing points to) — you don’t flag it.

Drafts tolerate "" (unwired) slots and unconfigured inputs so you can sketch. Publishing is strict — every slot must be wired to a real node and every required input filled. Validate before you publish with GET /v1/workflow/{workflowId}/validate-draft.

Plan anatomy

A complete, publishable plan for import a list → check email deliverability → add to a campaign:

{
  "name": "Verify emails from a list",
  "nodes": [
    {
      "id": "import",
      "kind": "BLOCK",
      "blockId": "SELECT_LEADS_FROM_LIST",
      "inputs": { "leadsGroupId": 55762 },
      "next": ["verify"]
    },
    { "id": "verify", "kind": "BLOCK", "blockId": "VERIFY_LEAD_EMAIL", "inputs": {}, "next": ["campaign"] },
    {
      "id": "campaign",
      "kind": "BLOCK",
      "blockId": "ADD_LEADS_TO_CAMPAIGN",
      "inputs": { "campaignId": 8842, "excludeContactedLeads": false },
      "next": [""]
    }
  ]
}

What makes it valid:

import is the entry node — nothing points to it, and it’s a selector block.
verify leaves inputs empty: its leadId input is the record flowing from the previous step, so you omit it. Only selector/standalone steps carry an explicit entity id.
campaign is terminal — its single next slot is "".
Every required input a publish needs is present (SELECT_LEADS_FROM_LIST.leadsGroupId, ADD_LEADS_TO_CAMPAIGN.campaignId).

Configured vs. flowing inputs. Entity ids (leadId, leadIds, companyId, companyIds, leadsGroupId, companyGroupId) are the record stream. Set them only on the entry/selector step; on every downstream step, omit them so the step receives the previous step’s output. GET /v1/workflows/blocks/ {blockId} spells out, per input, whether to configure it or let it flow.

Triggers

The optional trigger decides when a run starts. It defaults to MANUAL. All five types:

{ "type": "MANUAL", "entity": "LEAD" }

{
  "type": "SCHEDULE",
  "definition": "WEEKLY",
  "startsAt": "2026-07-27T08:00:00.000Z",
  "timezone": "Europe/Madrid"
}

{ "type": "WEBHOOK", "inputs": [{ "key": "contacts", "type": "leadIds", "label": "Contact IDs" }] }

{ "type": "API" }

{ "type": "INTENT", "signal": { "kind": "CONVERSATION_REPLY" } }

Type	When it fires	Notes
`MANUAL`	You start it (via `POST /v1/workflow/{id}/run` or the app)	Optional `entity` (`LEAD` / `COMPANY`) seeds the run.
`SCHEDULE`	On a recurrence	`definition` is `DAILY` / `WEEKLY` / `BI_WEEKLY` / `MONTHLY` / `QUARTERLY` / `YEARLY`; `startsAt` is ISO-8601. Attach an optional `recurrence` rule for finer control.
`WEBHOOK`	On delivery to the webhook URL	May declare typed `inputs` (max 20). The live URL is returned as `trigger.webhookUrl` on `GET`.
`API`	On an API-trigger call	For programmatic entry from your own systems.
`INTENT`	On an Enginy signal	`signal.kind` is one of `CONVERSATION_REPLY`, `WEBSITE_VISITOR`, `LEAD_CREATED`, `COMPANY_CREATED`, `LEAD_PROPERTY_UPDATED`, `COMPANY_PROPERTY_UPDATED`, `LEAD_ADDED_TO_LIST`, `COMPANY_ADDED_TO_LIST`.

A WEBHOOK trigger that declares inputs can’t be started with POST /v1/workflow/{id}/run — deliver its payload to the webhook URL instead. A webhook trigger with no declared inputs can still be run manually.

Drafts, published versions, and concurrency

Every workflow has one editable draft and, once published, one live published version.

Create / update always writes the draft. Nothing runs off a draft.
Publish snapshots the draft into a new immutable version (v1, v2, …) and reconciles the trigger (e.g. arms a schedule). Runs execute the published version.
Editing a published workflow is safe: the live version keeps running until you publish again.

Every GET /v1/workflow/{workflowId} returns a revision token (the workflow’s updatedAt, ISO-8601). Pass it back on PATCH …/draft and POST …/publish for optimistic concurrency:

Read

GET /v1/workflow/4021 → revision: "2026-07-19T10:15:30.000Z".

Write with the revision

PATCH /v1/workflow/4021/draft with "revision": "2026-07-19T10:15:30.000Z".

Handle 409

If someone else edited the draft in between, you get 409 Conflict — re-read, reconcile, retry with the new revision. Omit revision to force last-write-wins (not recommended for shared workflows). POST …/publish can also return 409 when another publish is mid-flight (its trigger is already being updated) — that one isn’t a revision conflict; just retry shortly.

Runtime input contract

A published workflow declares what a run needs via runtimeInputContract on GET /v1/workflow/{workflowId}. It has one of three statuses, and — for EMPTY and SUPPORTED — a ready-to-send exampleBody you can copy straight into the run call:

`status`	Meaning	How to run
`EMPTY`	The entry step is fully configured (e.g. a fixed list).	Call `run` with no `input`.
`SUPPORTED`	The entry step expects a record/list id at run time.	Pass `input` with one key from `acceptedKeys`.
`UNSUPPORTED`	The entry step can’t take a run-time id — it’s a `CONDITIONAL` or `KING`, or the plan declares more than one runtime input.	Not runnable via `run` — make the entry a selector/import block with a single `"__MANUAL_INPUT"`.

To make an entry step accept a run-time id, set its entity input to the sentinel "__MANUAL_INPUT":

{
  "id": "import",
  "kind": "BLOCK",
  "blockId": "SELECT_LEADS_FROM_LIST",
  "inputs": { "leadsGroupId": "__MANUAL_INPUT" },
  "next": ["verify"]
}

That yields a SUPPORTED contract whose acceptedKeys include the public alias (contactListId) and the raw key (leadsGroupId); the runner also accepts leadIds / leadId as a fallback. You’d then run it with, for example, { "input": { "contactListId": 55762 } } or { "input": { "leadIds": [101, 102] } }. The same sentinel drives WEBHOOK-triggered entry steps: declare an entity-typed webhook input (e.g. { "key": "contacts", "type": "leadIds" }) and set the entry step’s entity input to "__MANUAL_INPUT" — the run is seeded from the webhook payload at delivery time. Entity-typed webhook inputs are never wired with a { "webhookInputKey": … } binding; that binding style is only for non-entity fields (e.g. a profileUrl input filling a profileUrl field). For example, an IMPORT_COMPANY_FROM_PAYLOAD entry step with a declared { "key": "companyPayload", "type": "companyPayload" } webhook input must set inputs.companyPayload to { "webhookInputKey": "companyPayload" }. Put only configured values or bindings in node inputs — never copy the catalog’s key, type, required, valueShape, or resolution metadata into the plan. Also note the delivery credentials (trigger.webhookUrl + trigger.webhookSecret) are issued on publish — a draft has none yet.

Credits and limits

maxCreditsPerRun — an optional integer ceiling on credits a single run may spend. Enrichment and AI steps consume credits; a run stops once it hits the ceiling. Set it on create or PATCH …/draft.
60 nodes max per plan.
20 conditions max per CONDITIONAL node.

Access, scopes, and rate limits

Every request is authenticated with your API key in the x-api-key header (see Authentication). The base URL is https://openapi.enginy.ai.

	Scope	Rate limit (per key + endpoint)
Read endpoints	`WORKFLOWS_READ`	100 requests / 60s
Write endpoints	`WORKFLOWS_WRITE`	30 requests / 60s

Granting WORKFLOWS_WRITE implies WORKFLOWS_READ. A workflow is only visible to a key whose creating user can access it (workspace admins see all; other users see workflows they own or were granted access to).

Feature gating. The Workflows API requires the Advanced Workflows feature to be enabled for your workspace. If it isn’t, the endpoints return 403 Forbidden. Contact your Enginy admin to enable it.

Where to next

Endpoints

The interactive reference for every workflow endpoint, generated from the OpenAPI spec.

Cookbook

Four end-to-end builds you can copy and run.

MCP agent guide

The recommended tool flow for AI agents.

​What you work with

​Configure an import source from an audience

​The graph model

​Node kinds

​Wiring rules

​Plan anatomy

​Triggers

​Drafts, published versions, and concurrency

​Runtime input contract

​Credits and limits

​Access, scopes, and rate limits

​Where to next

Endpoints

Cookbook

MCP agent guide

What you work with

Configure an import source from an audience

The graph model

Node kinds

Wiring rules

Plan anatomy

Triggers

Drafts, published versions, and concurrency

Runtime input contract

Credits and limits

Access, scopes, and rate limits

Where to next