Cost Management

Gas Town burns tokens. A lot of tokens. At peak usage with 10+ polecats running simultaneously, you can expect to spend approximately $100/hour -- roughly 10x what a single Claude Code session costs. This guide covers monitoring, understanding, and optimizing your token spend.

Understanding Gas Town Costs

Where Tokens Go

Every agent session consumes tokens. The breakdown varies by workload, but a typical distribution looks like this:

Agent Type	% of Total Cost	Why
Polecats	60-70%	Writing code, running tests, iterating on implementations
Witnesses	8-12%	Patrol cycles, health checks, nudging stuck polecats
Refineries	5-10%	Merge conflict resolution, validation runs
Mayor	5-10%	Strategic coordination, convoy management
Deacon	3-5%	Health monitoring, lifecycle management
Dogs/Boot	1-3%	Infrastructure tasks, triage

Polecats dominate costs because they do the actual coding work -- reading files, reasoning about changes, running tests, and iterating until done.

Cost Factors

Factor	Impact
Number of polecats	Linear increase in cost
Task complexity	Complex tasks = more tokens per polecat
Codebase size	Larger codebases = more context tokens
Test suite runtime	Longer tests = longer sessions = more tokens
Merge conflicts	Conflicts spawn additional polecats for resolution
Patrol frequency	More frequent patrols = higher monitoring costs

Typical Cost Ranges

Usage Level	Polecats	Approx. Cost/Hour	Use Case
Minimal	1-2	$10-20	Focused work, single rig
Normal	3-5	$30-50	Active development, 1-2 rigs
Heavy	6-10	$50-100	Multi-rig parallel development
Peak	10-20+	$100-200+	Full fleet, aggressive timeline

Costs Add Up Fast

An 8-hour workday at "Normal" usage (3-5 polecats) costs $240-400. At "Peak" usage, a full day can exceed $1,000. Monitor continuously and adjust your polecat count based on budget constraints.

Monitoring Costs

`gt costs` Command

The primary cost monitoring tool:

# Current session costs
gt costs

# Today's costs
gt costs --today

# This week's costs
gt costs --week

# Breakdown by role
gt costs --by-role

# Breakdown by rig
gt costs --by-rig

# JSON output for external tools
gt costs --json

Sample Output

Gas Town Costs (today)

  Total tokens:  4,800,000 input / 1,620,000 output
  Est. cost:     $89.40

  By role:
    Polecats:    $61.30  (68.6%)  [12 sessions]
    Witnesses:   $9.80   (11.0%)  [3 sessions]
    Refineries:  $6.20   (6.9%)   [3 sessions]
    Mayor:       $7.10   (7.9%)   [1 session]
    Deacon:      $3.80   (4.3%)   [1 session]
    Other:       $1.20   (1.3%)

  By rig:
    myapp:       $52.10  (58.3%)
    api-server:  $24.80  (27.7%)
    docs:        $12.50  (14.0%)

  Hourly rate (last 3h avg): $32.40/hr

Real-Time Cost Monitoring

Keep a cost watch running alongside your feed:

# In terminal 1: activity feed
gt feed

# In terminal 2: cost updates every 5 minutes
watch -n 300 gt costs --today

Cost Optimization Strategies

Strategy 1: Fewer Polecats

The most direct way to reduce costs. Each polecat you remove saves its proportional token spend.

# Check how many polecats are running
gt polecat list

# Remove unnecessary polecats
gt polecat remove alpha --rig myproject

# Reduce the max polecat count via rig settings
gt rig settings set --rig myproject max_polecats 3

Quality Over Quantity

Three well-targeted polecats often outperform eight unfocused ones. Assign specific, well-defined tasks rather than vague directives. Specific tasks complete faster, consuming fewer tokens.

Strategy 2: Focused Convoys

Bundle related work tightly so polecats do not waste tokens on context-switching or redundant file reads.

# Instead of many small convoys:
gt convoy create "Fix A" gt-a1
gt convoy create "Fix B" gt-b2
gt convoy create "Fix C" gt-c3

# Create one focused convoy:
gt convoy create "Auth Module Fixes" gt-a1 gt-b2 gt-c3

When related beads are in the same convoy, the Mayor can assign them to fewer polecats that share context, reducing redundant file reading.

Strategy 3: Minimal Mode

Run Gas Town with reduced monitoring overhead for cost-sensitive workloads.

# Enable minimal mode
gt rig settings set minimal_mode true

Minimal mode:

Increases patrol cycle intervals (5 min to 15 min)
Reduces Witness verbosity
Skips non-essential health checks
Defers non-critical escalations

note

Minimal mode trades monitoring responsiveness for lower costs. Problems may take longer to detect. Use it for stable workloads where you are confident agents will not stall.

Strategy 4: Park Idle Rigs

Rigs you are not actively developing on still consume tokens through their Witnesses and Refineries:

# Park rigs you are not using today
gt rig park docs
gt rig park staging-env

# Unpark when needed
gt rig unpark docs

Strategy 5: Time-Box Sessions

Rather than running the fleet all day, work in focused sprints:

# Morning sprint (2 hours)
gt start --all
# ... intensive work ...
gt down

# Afternoon sprint (2 hours)
gt start --all
# ... intensive work ...
gt down

A 2-hour sprint at "Normal" usage costs $60-100, compared to $240-400 for an 8-hour continuous run.

Strategy 6: Reduce Context Size

Large codebases mean more input tokens per agent interaction. Reduce context costs by:

Using focused CLAUDE.md files that only include relevant context
Keeping project documentation concise
Using .gitignore patterns to exclude unnecessary files from agent view
Breaking monorepos into separate rigs so each polecat sees only its relevant code

Strategy 7: Optimize Merge Flow

Merge conflicts spawn additional polecats for resolution. Reduce conflicts by:

Assigning related files to the same polecat
Serializing work on highly-contended files via convoy dependencies
Running smaller, more frequent merges rather than large batches

Budget-Conscious Workflows

The "$50/day" Workflow

For teams with a $50/day budget:

# Use 2 polecats max
gt rig settings set max_polecats 2

# Enable minimal mode
gt rig settings set minimal_mode true

# Work in 2-hour focused sprints
gt start --all
# ... 2 hours of focused work ...
gt down

# Monitor costs
gt costs --today

The "$200/day" Workflow

For teams with a $200/day budget:

# Use up to 5 polecats
gt rig settings set max_polecats 5

# Normal monitoring
gt rig settings set minimal_mode false

# Work in 4-hour sprints with breaks
gt start --all
# ... 4 hours ...
gt down
# ... break ...
gt start --all
# ... 4 hours ...
gt down

# Monitor costs every hour
gt costs --today --by-role

The "Unlimited" Workflow

For teams where speed matters more than cost:

# Max polecats
gt rig settings set max_polecats 15

# Full monitoring
gt start --all

# Run all day, monitor hourly
watch -n 3600 gt costs --today

Cost Alerts

Set up automated cost alerts using the plugin system:

{
  "name": "cost-alert",
  "type": "schedule",
  "gate_type": "cron",
  "config": {
    "schedule": "0 * * * *",
    "command": "./check-costs.sh",
    "alert_threshold_hourly": 50,
    "alert_threshold_daily": 300
  }
}

Example check-costs.sh:

#!/bin/bash
DAILY_COST=$(gt costs --today --json | jq '.total_cost')

if (( $(echo "$DAILY_COST > 300" | bc -l) )); then
    gt escalate --severity high "Daily cost threshold exceeded: \$$DAILY_COST"
fi

Cost Comparison

To put Gas Town costs in perspective:

Resource	Cost/Hour	Equivalent
Gas Town (3 polecats)	~$30-50	1/3 of a senior developer hourly rate
Gas Town (10 polecats)	~$100	~1 senior developer hourly rate
Gas Town (20 polecats)	~$200	~2 senior developers hourly rate
Single Claude Code session	~$10	Baseline AI coding cost

The question is not "is this expensive?" but "is the throughput worth the cost?" Gas Town at 10 polecats can move 5-10x faster than a single developer, at roughly the same hourly cost.

Cost Tracking Best Practices

Check gt costs --today at least twice per day -- once in the morning and once before ending your session.
Set a daily budget and configure cost alerts to notify you when approaching it.
Review gt costs --by-role weekly to identify agents that are disproportionately expensive. A Witness consuming 20% of total cost may indicate a health monitoring loop.
Track cost per completed bead over time. This metric tells you whether you are getting more efficient: total_daily_cost / beads_closed_today.
Park unused rigs aggressively. Even idle Witnesses and Refineries consume tokens during patrol cycles.
Use gt down during breaks. Do not leave the fleet running during lunch or meetings unless you have active work in progress.

Understanding Gas Town Costs​

Where Tokens Go​

Cost Factors​

Typical Cost Ranges​

Monitoring Costs​

gt costs Command​

Sample Output​

Real-Time Cost Monitoring​

Cost Optimization Strategies​

Strategy 1: Fewer Polecats​

Strategy 2: Focused Convoys​

Strategy 3: Minimal Mode​

Strategy 4: Park Idle Rigs​

Strategy 5: Time-Box Sessions​

Strategy 6: Reduce Context Size​

Strategy 7: Optimize Merge Flow​

Budget-Conscious Workflows​

The "$50/day" Workflow​

The "$200/day" Workflow​

The "Unlimited" Workflow​

Cost Alerts​

Cost Comparison​

Cost Tracking Best Practices​