agenticstack-costs

Token and cost tracking harness for AgenticStack.

Accumulates prompt/completion/total tokens, call counts, and estimated USD spend per agent (and overall) by watching every LLM call through MONITOR-priority hooks — the same interception surface every marketplace plugin uses. Nothing is transformed; it only counts.

Install

pip install agenticstack-costs

Use

from agenticstack_costs import CostTracker

tracker = CostTracker(budget_usd=5.00).install()

# ... run any agents ...

print(tracker.report())
# {
#   "per_agent": {"Helper": {"calls": 3, "prompt_tokens": 2100,
#                            "completion_tokens": 640, "total_tokens": 2740,
#                            "cost_usd": 0.0021}},
#   "total": {..., "runs": 3},
#   "unpriced_calls": 0,
#   "over_budget": False,
# }

tracker.reset()      # zero the counters
tracker.uninstall()  # detach cleanly

Or as a plugin (CostsPlugin), activated by the plugin runtime; pass budget_usd through plugin config.

Pricing table

PRICES maps model-name prefixes to (input, output) USD per 1M tokens for gpt-4o, gpt-4o-mini, claude-3-5-sonnet, claude-3-5-haiku, gemini-1.5, and llama. The longest matching prefix wins (gpt-4o-mini-2024... prices as gpt-4o-mini, not gpt-4o). The numbers are estimates of public list prices — verify against your provider before relying on them. Models with no matching prefix cost 0.0 and increment the unpriced_calls counter.

Budgets are soft

budget_usd is tracking only: when the estimated total exceeds it, the tracker logs a warning and sets over_budget = True. It never blocks or stops anything — pair it with a guard plugin if you need hard enforcement.

Permissions

None — pure in-memory accounting.