Quickstart
Go from a blank canvas to your first full cost/latency estimate in under 15 minutes. No sign-up required.
Create a new workflow
Navigate to the Canvas view. Click "New Canvas" or press Ctrl+N. You'll land on a blank infinite canvas with a dot-grid background.
Give your canvas a descriptive name — it helps when comparing scenarios later.
Add nodes from the sidebar
Drag nodes from the left-hand palette onto the canvas. Start with a Start node, then add Agent nodes for your LLM calls, Tool nodes for external services, and a Finish node.
You need at least one Start node and one Finish node for estimation to work.
Configure each Agent
Double-click any Agent node to open its configuration modal. Select a Model Provider (OpenAI, Anthropic, Google, etc.), choose a Model, add context (system prompt or task description), and set the task type.
The more context you provide, the more accurate the token estimate. Paste a real system prompt if you have one.
Connect nodes with edges
Click a node's output handle and drag to another node's input handle to create an edge. Edges define execution order — data flows from source to target.
For parallel branches, connect a single source to multiple targets. The estimator will detect parallelism automatically.
Run the estimate
Click "Run Estimate" in the header. The estimate panel slides in with total token usage, cost per run, P95 latency, graph classification, and a per-node breakdown.
Estimation is pure computation — no API calls, no cost, under 10 ms.
Review the results
Check the Model Mix donut chart to see which models are most expensive. Review the Detailed Breakdown table for per-node input/output tokens, cost, and latency. Scroll to Scaling & Planning to project monthly costs.
Nodes consuming more than 50% of cost or latency are flagged in red — these are your bottlenecks.
Iterate and compare
Save the current scenario. Swap an expensive model for a cheaper alternative (e.g., GPT-4 → GPT-4o-mini). Run the estimate again. Open the Comparison Drawer to see side-by-side cost/latency differences.
Scenarios are free — create as many as you need to find the optimal cost-performance balance.
Try a sample workflow
Not sure where to start? Try this classic pattern:
This pipeline uses two different LLMs and a tool call. Running the estimate will show you exactly how cost and latency distribute across each step.
Try this workflowWhat to do next
Try swapping models on the same workflow and comparing estimates. For example, replace GPT-4o with GPT-4o-mini on the Research Agent to see how it affects cost versus latency.