Rungate vs Portkey.

# Portkey — virtual key model.
# Budget attaches to the key; the key is the unit of isolation.
# A workflow that spans multiple calls all use the same key —
# you can cap the key, but you can't cap the *workflow*.

from portkey_ai import Portkey

client = Portkey(
    api_key="PORTKEY_API_KEY",
    virtual_key="openai-vk-with-100usd-cap",  # budget on the key
)

# Each call in the workflow uses the same key.
# Portkey tracks cumulative spend per key, across all its calls.
# When the key hits its cap, the NEXT call fails.
r1 = client.chat.completions.create(...)   # ok
r2 = client.chat.completions.create(...)   # ok
# ... many calls later ...
rN = client.chat.completions.create(...)   # fails: key budget exceeded

# Rungate — run-level model.
# Budget attaches to the run (workflow). The agent token is a stable
# credential; the run is the unit of isolation.

from openai import OpenAI

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

# Every call in the workflow carries the same run ID.
# Rungate accrues budget against that run ID.
# When THIS workflow hits its cap, the next call in it fails.
# A separate workflow with 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 later ...
# HTTP 402 with context: { cumulative_spend, limit, run_id, rule }
rN = client.chat.completions.create(
    ..., extra_headers={"x-rungate-run-id": RUN},
)