Skip to main content
Forge

Platform

The control plane underneath every workflow

Eight capabilities make up the Forge platform. They are designed to be used together, but each one is addressable on its own through the API — so you can adopt the platform incrementally instead of all at once.

AI workflows

Intelligence steps with typed inputs, structured outputs and evaluation attached from day one.

  • Prompt and schema versioned together as one revision
  • Retrieval over indexed sources with citation output
  • Confidence routing to a human review queue
  • Golden-set evaluation on every publish

Automation engine

Durable execution for processes measured in days rather than milliseconds.

  • Checkpointed steps that resume after a restart
  • Exponential backoff with a dead-letter queue
  • Parallel fan-out and conditional joins
  • Idempotency keys on every external write

Data layer

Governed access to the records a workflow needs, without copying your warehouse.

  • Read-through connectors to warehouse and object storage
  • Field-level masking for sensitive attributes
  • Configurable retention windows per workspace
  • Lineage from a written record back to the run that produced it

Integrations

Typed connectors that fail loudly when an upstream contract changes.

  • HTTP, queue, warehouse and object-store connectors
  • Schema drift detection surfaced as a failing check
  • Per-connection health and error-rate monitoring
  • Sandbox connections for safe development

APIs & SDKs

Everything in the console is an API call, documented and versioned.

  • OpenAPI specification with generated TypeScript and Python clients
  • Versioned endpoints with a published deprecation window
  • Cursor pagination and consistent error envelopes
  • Per-key rate limits with headers you can act on

Events & webhooks

Outbound delivery you can trust, and a replay button for when you cannot.

  • Signed payloads with a rotating per-endpoint secret
  • At-least-once delivery with automatic retry
  • Replay any event from the last retention window
  • Delivery log with response codes and timing

Permissions

Roles that map to how the team is actually organised.

  • Role-based access scoped to workspace, project or workflow
  • SSO with SCIM provisioning and de-provisioning
  • Named approvers required before a publish takes effect
  • Scoped API credentials with independent rotation

Analytics

Cost, latency, throughput and accuracy — attributed to the step that caused them.

  • Per-step cost attribution including model spend
  • Latency percentiles rather than averages
  • Accuracy trend against human corrections
  • Warehouse export for your own reporting

Runtime

How a run actually executes

A Forge run is a state machine with a durable log. That single design decision is what makes retries safe, audits complete and long-running processes possible.

  1. 01

    Accept

    Input is validated against the workflow schema and assigned a run identifier.

  2. 02

    Plan

    The engine resolves the step graph for the pinned workflow revision.

  3. 03

    Execute

    Steps run with checkpoints; failures retry with backoff instead of restarting the run.

  4. 04

    Escalate

    Low confidence or a hard failure routes the run to the assigned review queue.

  5. 05

    Commit

    Writes happen behind idempotency keys, then the audit event is sealed.

forge / api console

POST /v1/workflows/invoice-intake/runsauthorization: Bearer sk_live_••••••••idempotency-key: inv-2026-08-17-0091 200 OK   142ms{  "runId": "run_9f21ac",  "status": "completed",  "steps": 6,  "reviewed": false}
Demo interface: an API console showing a workflow run request and its structured response.

Illustrative platform limits

Max steps per workflow
250
Max run duration
30 days
Payload size per step
10 MB
Event retention
90 days
Concurrent runs
Plan dependent
API rate limit
600 req / min

Demo figures shown to make the template concrete. Replace with your own numbers.

Adjacent

Where to go next

The product overview shows these capabilities in the context of a real workflow. The security page covers the control model in detail.

Get started

Adopt one capability, then the rest

Most teams start with a single workflow and one connector. The platform is designed so that first step does not become a rewrite later.