token-budget-py 0.1.0

Thread-safe shared token + USD budget for concurrent LLM tasks. Raises BudgetExceeded on push past cap. Reserve/commit two-phase API for fan-out workloads. Zero runtime deps.

token-budget-py

Thread-safe shared token + USD budget for concurrent LLM tasks.

Fan-out workloads — agents, parallel summarizers, batch evals — race many tasks to consume from one shared budget. This library is a small, zero-dependency counter with two axes (tokens, USD) that returns BudgetExceeded when a record would push past a configured cap.

Sibling to the Rust crate token-budget-pool.

Install

pip install token-budget-py

Use

from token_budget import BudgetPool, BudgetExceeded

pool = BudgetPool(token_cap=1_000_000, usd_cap=10.0)

try:
    pool.record(tokens=1200, usd=0.0036)
except BudgetExceeded as e:
    # tell this worker to skip
    print(f"out of budget: {e}")

Two-phase commit (reserve before the LLM call, commit the actual usage):

with pool.reserve(tokens=2000, usd=0.012) as r:
    result = call_llm(prompt)
    r.commit(tokens=result.usage.total_tokens, usd=result.cost_usd)

If the with block exits without r.commit() (e.g. the LLM call raised), the reservation is auto-released — no orphaned slots.

Either axis is optional:

only_tokens = BudgetPool(token_cap=500_000)        # USD unbounded
only_usd    = BudgetPool(usd_cap=5.0)              # tokens unbounded
unbounded   = BudgetPool()                         # both unbounded (counter only)

Atomic read of current state:

snap = pool.snapshot()
snap.tokens_used         # 1200
snap.usd_remaining       # 9.9964
snap.tokens_remaining    # 998800 (cap - used - reserved)

What it does NOT do

• No async runtime lock-in. Works under asyncio, trio, threads, sync. The internal lock is a plain threading.Lock (held only for the microseconds of a counter update).
• No HTTP. Doesn't talk to any LLM provider.
• No cost calculation. Wrap a cost calculator that returns USD per call and feed the result into record. (See claude-cost, openai-cost, gemini-cost, bedrock-cost on crates.io for Rust cost calculators with the same authorship.)
• No persistence. Counts live in process. For multi-process / multi-host budgets, wrap a Redis or DB increment instead.
• No automatic rollover. Call pool.reset() from your own cron / time loop if you want a periodic window.

License

MIT