Skip to content

arieradle/shekel

Repository files navigation

🪙 shekel

Stop your AI agent from bankrupting you.

PyPI version Python versions CI codecov Downloads License: MIT GitHub Stars Docs

with budget(max_usd=5.00):
    run_my_agent()       # hard stop at $5. no SDK changes. no config. just works.
shekel run agent.py --budget 5   # or enforce without touching code at all

I woke up to a $47 OpenAI bill from a LangGraph agent that spent the night retrying a failed tool call. I built shekel so you don't have to learn that lesson yourself.


Install

pip install shekel[openai]       # OpenAI
pip install shekel[anthropic]    # Anthropic
pip install shekel[all]          # OpenAI + Anthropic + LiteLLM + Gemini + HuggingFace
pip install shekel[cli]          # shekel run — enforce budgets without touching code

Works with everything

If it calls OpenAI or Anthropic under the hood, shekel sees it — zero integration code needed.

Provider Framework
OpenAI · Anthropic · Gemini LangChain · LangGraph Auto-patched
HuggingFace · LiteLLM · Groq CrewAI · OpenAI Agents SDK Auto-patched
MCP · AutoGen · LlamaIndex Any custom wrapper Auto-patched

Every pattern you'll actually use

Hard cap — the one that saves you

from shekel import budget

with budget(max_usd=5.00):
    run_my_agent()
# raises BudgetExceededError the moment spend crosses $5

No wrapping your OpenAI client. No decorators. No SDK replacement. shekel monkey-patches the provider on context entry and restores it on exit. Your existing code runs unchanged.

Warn before the limit hits

with budget(max_usd=5.00, warn_at=0.8) as b:
    run_my_agent()
# logs a warning at $4.00, raises at $5.00

Track spend without enforcing

with budget() as b:
    run_my_agent()
print(f"that cost ${b.spent:.4f}")

Switch to a cheaper model instead of crashing

with budget(max_usd=1.00, fallback={"at_pct": 0.8, "model": "gpt-4o-mini"}) as b:
    run_my_agent()
# switches gpt-4o → gpt-4o-mini at $0.80, hard stops at $1.00

Cap tool calls — stop the infinite search loop

from shekel import tool

@tool(price=0.01)               # charge $0.01 per call + count toward the cap
def web_search(query: str) -> str: ...

@tool                           # free — just count calls
def read_file(path: str) -> str: ...

with budget(max_usd=5.00, max_tool_calls=20) as b:
    run_my_agent()
# ToolBudgetExceededError on call 21 — before the tool runs
print(b.summary())              # LLM spend + tool spend broken out by tool name

Auto-intercepted with zero config: LangChain, MCP, CrewAI, OpenAI Agents SDK.

Per-stage budget control

with budget(max_usd=10.00, name="pipeline") as pipeline:
    with budget(max_usd=2.00, name="research"):
        results = search_web(query)        # capped at $2

    with budget(max_usd=5.00, name="analysis"):
        report = analyze(results)          # capped at $5

print(pipeline.tree())
# pipeline: $4.80 / $10.00
#   research:  $1.20 / $2.00
#   analysis:  $3.60 / $5.00

Children auto-cap to the parent's remaining balance. b.tree() gives you a live visual breakdown.

LangGraph — per-node circuit breaking

with budget(max_usd=10.00, name="graph") as b:
    b.node("fetch_data", max_usd=0.50)   # NodeBudgetExceededError before node runs
    b.node("summarize",  max_usd=1.00)

    app = graph.compile()
    app.invoke({"query": "..."})

print(b.tree())
# graph: $0.84 / $10.00
#   [node] fetch_data: $0.12 / $0.50  (24%)
#   [node] summarize:  $0.72 / $1.00  (72%)

Shekel patches StateGraph.add_node() transparently — no graph changes needed.

LangChain — per-chain circuit breaking

with budget(max_usd=5.00, name="pipeline") as b:
    b.chain("retriever",  max_usd=0.20)   # ChainBudgetExceededError before chain runs
    b.chain("summarizer", max_usd=1.00)

    retriever_chain.invoke({"query": "..."})
    summarizer_chain.invoke({"doc": "..."})

Shekel patches Runnable._call_with_config and RunnableSequence.invoke — zero changes to your chains.

CrewAI — per-agent and per-task circuit breaking

from shekel.exceptions import AgentBudgetExceededError, TaskBudgetExceededError

try:
    with budget(max_usd=5.00, name="crew") as b:
        b.agent(researcher.role,       max_usd=2.00)  # use agent.role directly
        b.agent(writer.role,           max_usd=1.00)
        b.task(research_task.name,     max_usd=1.50)  # use task.name directly
        b.task(write_task.name,        max_usd=0.80)
        crew.kickoff(inputs={"topic": "AI"})
except TaskBudgetExceededError as e:
    print(f"Task '{e.task_name}' over budget: ${e.spent:.4f} / ${e.limit:.2f}")
except AgentBudgetExceededError as e:
    print(f"Agent '{e.agent_name}' over budget")

print(b.tree())
# crew: $2.84 / $5.00
#   [agent] Senior Researcher: $1.92 / $2.00  (96.0%)
#   [agent] Content Writer:    $0.92 / $1.00  (92.0%)
#   [task]  research:          $1.92 / $1.50  (128.0%)
#   [task]  write:             $0.92 / $0.80  (115.0%)

Shekel patches Agent.execute_task transparently. Gate order: task cap → agent cap → global (most specific first).

Distributed budgets — enforce across multiple processes

from shekel.backends.redis import RedisBackend

backend = RedisBackend()   # reads REDIS_URL from env; fail-closed by default

with budget("$5/hr + 100 calls/hr", name="api-tier", backend=backend) as b:
    response = client.chat.completions.create(...)
# Atomic Lua-script enforcement — one Redis round-trip per call
# BudgetConfigMismatchError if the same name is reused with different limits

Works with AsyncRedisBackend for async workflows. Circuit breaker built in — configurable threshold + cooldown. Fail-open or fail-closed.

Rolling-window rate limits

with budget("$5/hr", name="api-tier") as b:
    response = await client.chat.completions.create(...)
# BudgetExceededError carries retry_after so callers know when the window resets

Multi-cap: budget("$5/hr + 100 calls/hr") — USD and call-count windows are independent.

Accumulate across sessions

session = budget(max_usd=20.00, name="session")

with session: run_step_1()   # $3.20
with session: run_step_2()   # $8.10
with session: run_step_3()   # raises at $20

print(f"total: ${session.spent:.2f}")

Enforce from the CLI — zero code changes

Don't want to touch the code at all? Don't.

pip install shekel[cli]

shekel run agent.py --budget 5
# exit 0 = under budget  |  exit 1 = budget exceeded  ← CI-friendly

Drop it into any pipeline:

# Shell / cron / Docker
AGENT_BUDGET_USD=5 shekel run agent.py

# GitHub Actions
- run: shekel run agent.py --budget 5

# Docker — operator sets budget at runtime, no rebuild needed
ENTRYPOINT ["shekel", "run", "agent.py"]
# docker run -e AGENT_BUDGET_USD=5 my-agent-image

Key flags:

--budget 5          # hard stop in USD
--warn-at 0.8       # log warning at 80%, hard stop at 100%
--max-llm-calls 20  # cap by call count instead of spend
--max-tool-calls 50 # cap agent tool calls
--warn-only         # log but never exit 1  (soft guardrail)
--dry-run           # track costs, no enforcement
--output json       # machine-readable spend summary for log pipelines
--budget-file shekel.toml  # load limits from config file

What the spend summary looks like

with budget(max_usd=5.00) as b:
    run_my_agent()

print(b.summary())
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
shekel spend summary
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Total: $1.2450 / $5.00 (25%)

gpt-4o:       $1.1320  (5 calls)
  Input:  45.2k tokens → $0.1130
  Output: 11.3k tokens → $1.1320

Tool spend:   $0.1130  (9 tool calls)
  web_search  $0.090  (9 calls)  [langchain]
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Or machine-readable:

shekel run agent.py --budget 5 --output json
# {"spent": 1.245, "limit": 5.0, "calls": 5, "tool_calls": 9, "status": "ok", "model": "gpt-4o"}

The decorator

from shekel import with_budget

@with_budget(max_usd=0.10)
def summarize(text: str) -> str:
    return client.chat.completions.create(...).choices[0].message.content
# budget enforced independently on every call

How it works

shekel monkey-patches openai.chat.completions.create and anthropic.messages.create on __enter__ and restores originals on __exit__. Spend is tracked in a ContextVar — concurrent agents in the same process never share state. Nested with budget() blocks form a tree; child spend rolls up automatically.

No background threads. No external services. No API keys. Nothing leaves your machine.


Observability

  • Langfuse — cost streaming, circuit-break events, budget hierarchy in Langfuse spans
  • OpenTelemetry — 9 instruments: shekel.llm.cost_usd, shekel.budget.utilization, shekel.budget.spend_rate, shekel.tool.calls_total, and more
from shekel.otel import ShekelMeter
meter = ShekelMeter()  # attaches to global MeterProvider; silent no-op if OTel absent

Supported models

Built-in pricing for GPT-4o, GPT-4o-mini, o1, o3, Claude 3.5/3/3.7 Sonnet, Claude 3 Haiku/Opus, Gemini 2.0/2.5 Flash/Pro, and more.

pip install shekel[all-models]   # 400+ models via tokencost
shekel models                    # list all bundled models and pricing
shekel estimate --model gpt-4o --input-tokens 1000 --output-tokens 500

API quick reference

budget(
    max_usd=5.00,           # hard USD cap
    warn_at=0.8,            # warn at 80%
    max_llm_calls=50,       # cap by call count
    max_tool_calls=100,     # cap tool dispatches
    tool_prices={"web_search": 0.01},  # charge per tool
    fallback={"at_pct": 0.8, "model": "gpt-4o-mini"},  # switch instead of crash
    name="my-agent",        # required for nesting + temporal budgets
    backend=RedisBackend(), # distributed enforcement across processes
)

budget("$5/hr + 100 calls/hr", name="api-tier")  # multi-cap rolling-window

Component caps — all chainable, all raise before the component executes:

b.node("fetch_data", max_usd=0.50)   # LangGraph node  → NodeBudgetExceededError
b.chain("retriever", max_usd=0.20)   # LangChain chain → ChainBudgetExceededError
b.agent("researcher", max_usd=1.00)  # CrewAI agent    → AgentBudgetExceededError
b.task("summarize", max_usd=0.50)    # CrewAI task     → TaskBudgetExceededError

Exceptions — all subclass BudgetExceededError, so one except catches everything:

Exception Raised when Key fields
BudgetExceededError Global cap hit spent, limit, model, retry_after
NodeBudgetExceededError LangGraph node cap hit node_name, spent, limit
AgentBudgetExceededError CrewAI agent cap hit agent_name, spent, limit
TaskBudgetExceededError CrewAI task cap hit task_name, spent, limit
ChainBudgetExceededError LangChain chain cap hit chain_name, spent, limit
ToolBudgetExceededError Tool call cap hit tool_name, calls_used, calls_limit
BudgetConfigMismatchError Redis name reused with different limits

Security

Every PR and push to main runs CodeQL, Trivy, Bandit, and pip-audit. See the Security tab for results.


Documentation

arieradle.github.io/shekel


Contributing

See CONTRIBUTING.md. PRs welcome — especially new framework adapters.

License

MIT

About

LLM budget control and cost governance for AI agents. Python library for token budgets, usage limits and guardrails for OpenAI, Anthropic, LangChain, LangGraph and agentic systems.

Topics

Resources

License

Contributing

Stars

Watchers

Forks

Contributors

Languages