Rungate vs LiteLLM.

# LiteLLM — virtual key with per-key budget.
# The key is the unit of spend tracking + isolation.
# The proxy blocks requests synchronously when the key's cap is hit.

import openai

client = openai.OpenAI(
    base_url="https://your-litellm-proxy/v1",
    api_key="sk-your-virtual-key-with-1usd-cap",  # budget on the key
)

# All calls in this workflow use the same virtual key.
# LiteLLM accrues cost-per-token into the key's running total.
# When the key hits max_budget, the NEXT call fails.
r1 = client.chat.completions.create(...)   # ok
# ... workflow calls ...
# HTTP 401 "Authentication Error, ExceededTokenBudget: ..."
rN = client.chat.completions.create(...)
# Note: LiteLLM returns 401 for budget-exceeded, which is
# semantically an auth failure. HTTP 402 Payment Required would
# be more precise for spending-cap violations.

# Rungate — run-level model.
# The agent token is a stable credential, not a budget carrier.
# Budget attaches to the run (workflow) via policy.

from openai import OpenAI

client = OpenAI(
    base_url="https://api.rungate.dev/v1",
    api_key="rg_agt_your_agent_token_here",
)

# Every call in THIS workflow carries the run ID.
# Rungate accrues spend against that run.
# Tomorrow's workflow under the same token has its own budget.
RUN = "run_workflow_abc123"
r1 = client.chat.completions.create(
    ..., extra_headers={"x-rungate-run-id": RUN, "x-rungate-policy": "prod-v4"},
)
# ... many calls + tool invocations + retries later ...
# HTTP 402 Payment Required with context: { cumulative_spend,
# limit_usd, run_id, rule, steps_completed }.
# Agent parses context, stops dispatching on this run, starts a
# new run with a fresh budget if appropriate.