Your autonomous agents are burning money while you sleep

OpenClaw runs agents 24/7 — heartbeats, brain ticks, nightly builds. Most of those calls are simple status checks hitting opus-4. NadirClaw routes them to flash automatically. Save 50-70%.

Set Up in 5 Minutes See the Math

The autonomous agent cost problem

OpenClaw makes it easy to run agents around the clock. That's the point. But it also means your bill never stops growing.

Heartbeats hit your expensive model

Every agent checks in 6x/day. That's "I'm alive, nothing to report" routed to opus-4 at $15/1M input tokens. It should cost $0.001, not $0.15.

Brain ticks are mostly idle

Hourly brain ticks scan for work. 90% of the time: "nothing new." But each tick still burns a full Opus call because that's the configured model.

It scales linearly with agents

5 agents × 6 heartbeats × 24 brain ticks = 144 calls/day before any real work happens. At Opus prices, that's $20+/day in overhead alone.

Setup: 5 minutes, 2 steps

NadirClaw runs locally as a proxy. Point OpenClaw at it. Done.

STEP 1

Install and start NadirClaw

NadirClaw runs on port 8856 and proxies to any LLM provider. The classifier runs locally — no data leaves your machine.

STEP 2

Point OpenClaw at the proxy

One config change. OpenClaw sends all LLM calls through NadirClaw, which classifies complexity and routes to the cheapest model that can handle it.

The config

Two files. Copy, paste, done.

Terminal — Install NadirClaw

pip install nadirclaw
nadirclaw serve --port 8856

~/.openclaw/config.yaml — Point OpenClaw at NadirClaw

# OpenClaw global config
llm:
  base_url: "http://localhost:8856"
  default_model: "claude-opus-4"
  api_key: "${ANTHROPIC_API_KEY}"

nadirclaw.yaml — Route config (optional, sensible defaults work)

# NadirClaw routing rules
classifier:
  model: "local"           # runs on-device, no API cost
  threshold: 0.6           # complexity score cutoff

routes:
  simple:
    - gemini-2.0-flash     # $0.10/$0.40 per 1M tokens
    - claude-haiku-3.5     # $0.80/$4.00 per 1M tokens
  complex:
    - claude-opus-4         # $15/$75 per 1M tokens
    - claude-sonnet-4       # $3/$15 per 1M tokens

fallback: claude-sonnet-4   # if classifier is unsure

Real example: 5-agent OpenClaw setup

A typical production setup. Let's do the math.

You're running 5 OpenClaw agents: engineering lead, PM, QA, DevOps, and a research agent. Each runs 6 dispatches/day, brain ticks every hour, plus nightly builds. Here's what a typical day looks like:

Task Type	Daily Calls	Complexity	Without NadirClaw	With NadirClaw
Heartbeats (5 agents × 6/day)	30	Simple	$4.50 (Opus)	$0.03 (Flash)
Brain ticks (5 agents × 24/day)	120	90% Simple	$18.00 (Opus)	$2.52 (108 Flash + 12 Opus)
Dispatch work (coding, reviews)	30	Complex	$13.50 (Opus)	$13.50 (Opus — stays)
Nightly builds (5 agents)	5	Mixed	$3.75 (Opus)	$1.50 (3 Flash + 2 Opus)
Daily Total	185		$39.75	$17.55

56% saved — $22.20/day → $666/month

Complex work stays on Opus. Only the simple stuff moves to Flash.
Quality where it matters. Savings everywhere else.

What it looks like running

NadirClaw classifying OpenClaw agent traffic in real time.

nadirclaw serve --port 8856

✓ Classifier ready (local model)

✓ Listening on localhost:8856

✓ OpenClaw agents detected: 5 connected

forge simple "session_status: idle, no tasks" → flash-2.0 $0.001

scout simple "brain tick: checking BACKLOG.md..." → flash-2.0 $0.001

forge complex "refactor auth middleware to support..." → opus-4 $0.450

sentinel simple "nightly build: all tests passing" → flash-2.0 $0.001

clawy complex "draft Q1 roadmap considering..." → opus-4 $0.380

Start saving now

Takes less time than reading this page did.

Get NadirClaw Read the Docs