Imported from jeremylongshore/tons-of-skills-marketplace (skills/.curated/lindy-cost-tuning/SKILL.md). Install upstream with npx skills add jeremylongshore/tons-of-skills-marketplace --skill lindy-cost-tuning. Copyright stays with the author (MIT).

Lindy Cost Tuning

Overview

Lindy uses a credit-based pricing model. Every task costs credits based on model size, step count, premium actions, and duration. Cost tuning targets: model right-sizing, agent consolidation, trigger optimization, and credit monitoring.

Prerequisites

• Lindy workspace with billing access
• Multiple active agents to evaluate
• Dashboard access to review per-agent task history

Credit Cost Reference

Plan Costs

Instructions

Step 1: Audit Agent Credit Consumption

For each active agent, collect:

1. Task count (last 30 days) — from Tasks tab
2. Average credits per task — total credits / task count
3. Model used — from agent settings
4. Trigger frequency — how often the agent fires

Create a cost audit table:

Step 2: Right-Size Models

The highest-impact optimization. For each agent, ask:

"Does this task actually need GPT-4/Claude, or would Gemini Flash work?"

Test the downgrade: Run 10 tasks with the smaller model. Compare output quality. Most classification, routing, and extraction tasks work identically on smaller models.

Step 3: Consolidate Redundant Agents

Multiple single-purpose agents cost more than one multi-purpose agent:

Before (5 agents, 5 minimum credits per run):

Agent 1: Classify billing emails
Agent 2: Classify technical emails
Agent 3: Classify general emails
Agent 4: Draft billing responses
Agent 5: Draft technical responses

After (1 agent, 1 minimum credit per run):

Support Agent: Classify email → Condition (billing/technical/general)
  → Draft appropriate response → Send

Cost impact: Reducing from 5 agents to 1 saves minimum-credit overhead and simplifies management.

Step 4: Optimize Trigger Frequency

Credits are consumed every time a trigger fires. Reduce unnecessary triggers:

Email Received:

Before: Trigger on ALL emails (300/day) = 300 tasks
After:  Filter: label "support" AND NOT from "noreply@" (40/day) = 40 tasks
Savings: 87% fewer tasks

Schedule trigger:

Before: Every 15 minutes (96/day)
After:  Every 2 hours (12/day)
Question: Does this agent really need to run every 15 minutes?

Slack trigger:

Before: Any message in #general (200/day)
After:  Messages containing "@support-bot" (10/day)
Savings: 95% fewer tasks

Step 5: Reduce Steps Per Task

Each action in a workflow costs credits. Eliminate unnecessary steps:

• Combine multiple LLM calls into one (see lindy-performance-tuning)
• Use Set Manually instead of AI Prompt for known values
• Remove debug/logging steps in production
• Simplify condition branches

Step 6: Optimize Knowledge Base Usage

KB search costs credits per query. Optimize:

• Reduce Max Results from 10 to 4 (sufficient for most queries)
• Use specific query instructions to get relevant results in one search
• For small datasets (<100 entries), consider putting data directly in the prompt

Step 7: Budget Monitoring Setup

1. Check credit usage weekly in Settings > Billing
2. Set internal alerts for high-consumption agents:
   • 50% of budget: Warning — review usage
   • 80% of budget: Alert — optimize or upgrade
   • 95% of budget: Critical — pause non-essential agents

Step 8: Deactivate Idle Agents

Review agents monthly:

• No tasks in 30 days → Pause the agent
• No tasks in 90 days → Delete or archive
• Lindy only charges for active agent execution, not idle agents

Monthly Cost Optimization Checklist

• Review per-agent credit consumption
• Identify agents using large models for simple tasks
• Check for redundant agents that could be consolidated
• Review trigger filter effectiveness
• Remove unused integrations from agents
• Verify no loops or runaway agent steps
• Compare actual spend to budget

Error Handling

Output

Produce a cost-control report showing the workload volume, credit-consuming steps, monthly budget, proposed optimization, and the expected service-quality effect. Measure before and after the change rather than inferring savings from agent configuration alone.

Examples

For a high-volume enrichment agent, move deterministic validation ahead of the credit-consuming enrichment step and test it against a representative week of events. Compare completed useful actions, skipped invalid inputs, and credits per successful outcome before deploying the new ordering.

Resources

• Lindy Pricing
• Lindy Documentation

Next Steps

Proceed to lindy-reference-architecture for production architecture patterns.