Your workflows are still
waiting for you to click Run

Quick-start via CLI or API

Orchestrate ships a single CLI command that scaffolds your first agent with model binding, retry policy, and webhook endpoint pre-configured.

$ npm install -g @orchestrate/cli
$ orchestrate init --model gpt-4o --name invoice-parser
✓ Agent scaffolded in ./agents/invoice-parser
✓ Webhook endpoint: https://run.orchestrate.io/agt-001
✓ Retry policy: 3x exponential backoff
$ orchestrate deploy
→ Agent live in 4.2s

Model registry — 40+ providers

Orchestrate maintains a live model registry covering OpenAI, Anthropic, Google Gemini, Mistral, Cohere, and 35+ open-source models via Ollama and HuggingFace endpoints. Model routing can be rule-based (cost, latency, capability) or AI-selected per task.

// model_config.yaml
model:
  primary: gpt-4o
  fallback: claude-3-5-sonnet
  routing:
    cost_ceiling: $0.02/task
    latency_sla: 3s
    auto_failover: true

DAG-based execution graph

Orchestrate models workflows as directed acyclic graphs (DAGs). Each node can branch to N downstream nodes based on output conditions, model confidence scores, or arbitrary custom logic. Cycles are supported for retry loops.

// workflow.json (excerpt)
{
  "nodes": {
    "classify": { "model": "gpt-4o-mini", "next": ["approve", "escalate", "reject"] },
    "approve": { "action": "crm.updateStatus", "next": ["notify"] },
    "escalate": { "action": "slack.alert", "condition": "confidence < 0.7" }
  }
}

Three-layer error recovery

Layer 1: Per-step retry with exponential backoff (configurable). Layer 2: Fallback model routing when primary LLM fails or exceeds latency SLA. Layer 3: Dead-letter queue with human escalation webhook for unrecoverable failures.

retry_policy:
  max_attempts: 3
  backoff: exponential
  initial_delay: 500ms
  max_delay: 30s
  on_exhausted:
    action: dead_letter_queue
    notify: ops@yourco.com
    include: [trace_id, inputs, error]

Immutable execution traces

Every agent execution produces a cryptographically signed trace log including: model inputs/outputs, tool calls, latency breakdown per step, token counts, cost attribution, and operator identity. Logs are queryable via API and exportable to S3, Datadog, or Splunk.

GET /v1/traces/agt-001/exec-2026022705104
{
  "trace_id": "tr_8xK2mN9pQ",
  "agent": "invoice-parser",
  "duration_ms": 1243,
  "steps": [
    { "step": "extract", "model": "gpt-4o", "tokens": 847, "cost": 0.0042 },
    { "step": "validate", "tool": "schema_check", "passed": true },
    { "step": "post", "action": "crm.create", "status": 201 }
  ],
  "signed": "sha256:a3f9..."
}

Flat per-execution pricing — no task multiplication

Zapier bills per task, meaning a 10-step workflow processing 100 records = 1,000 tasks billed. Orchestrate bills per agent execution regardless of internal step count or record volume. A 50-step workflow processing 10,000 records = 1 execution billed.

Rate limit specifications

Orchestrate rate limits are enforced at the organization level, not per-workflow. Burst allowance is 3× sustained rate for up to 60 seconds. Limits scale automatically with plan tier and can be raised on request with 24h SLA.

// Rate limit headers on every response
X-RateLimit-Limit: 10000
X-RateLimit-Remaining: 9847
X-RateLimit-Reset: 1740632400
X-Burst-Remaining: 29847
Retry-After: (only on 429)