llmkit-sdk 0.1.7

AI API gateway SDK with cost tracking, budget enforcement, and multi-provider routing for LLM applications

Track what your AI agents cost. One line of code.

---

Cost tracking for LLM APIs. Works with OpenAI, Anthropic, Gemini, Groq, Mistral, Together, and Cohere SDKs. 730+ models priced. Zero config, zero account needed for local tracking.

pip install llmkit-sdk

Zero-config cost tracking

Wrap any OpenAI-compatible client. Costs are estimated locally from a bundled pricing table - no proxy, no account, no network calls:

from llmkit import tracked
from openai import OpenAI

client = OpenAI(http_client=tracked())
res = client.chat.completions.create(
    model="gpt-4.1",
    messages=[{"role": "user", "content": "explain CQRS"}],
)
# costs calculated automatically from response usage data

Works the same with Anthropic, Gemini, Groq, Mistral, Together, and Cohere:

from anthropic import Anthropic

client = Anthropic(http_client=tracked())
msg = client.messages.create(
    model="claude-sonnet-4-20250514",
    max_tokens=1024,
    messages=[{"role": "user", "content": "explain event sourcing"}],
)

Collect costs

costs = []
client = OpenAI(http_client=tracked(on_cost=costs.append))

# ... run your agent ...

total = sum(c.total_cost for c in costs if c.total_cost)
print(f"Agent run cost: ${total:.4f}")

Estimate from any response

from llmkit import estimate_cost

cost = estimate_cost(response)
print(f"~${cost.total_cost:.6f}")

How it compares

Feature	llmkit-sdk	tokencost	litellm
Zero-config tracking	yes (httpx transport)	no (manual call)	no (callback setup)
Works with existing SDK code	yes (drop-in)	no (separate function)	yes (but requires litellm wrapper)
Local estimation (no proxy)	yes	yes	no
Budget enforcement	yes (via proxy)	no	yes (but 9+ bypass bugs)
Streaming cost tracking	yes	no	yes
Session grouping	yes	no	no
Models priced	730+	400+	100+
Install size	~200KB	~50KB	~50MB

Session tracking

Group costs by agent run:

from llmkit import LLMKit

client = LLMKit(api_key="llmk_...")
agent = client.session()

for task in tasks:
    completion, cost = agent.chat(
        model="gpt-4.1",
        messages=[{"role": "user", "content": task}],
    )

print(f"Session: ${agent.stats.total_cost:.4f} across {agent.stats.request_count} requests")

Streaming

stream = client.chat_stream(
    model="claude-sonnet-4-20250514",
    messages=[{"role": "user", "content": "write a haiku"}],
)
for chunk in stream:
    print(chunk.choices[0].delta.content or "", end="")

print(f"\nCost: ${stream.cost.total_cost:.6f}")

Proxy mode (budget enforcement)

Route through the LLMKit proxy for hard budget limits, per-key rate limiting, and dashboard analytics:

client = LLMKit(api_key="llmk_...")
completion, cost = client.chat(
    model="gpt-4.1",
    messages=[{"role": "user", "content": "hello"}],
)
print(f"${cost.total_cost:.4f} via {cost.provider}")

Set a $10 daily budget in the dashboard. When it's hit, requests get a 402 - not a log message, an actual block. No more runaway agents.

Async

from llmkit import AsyncLLMKit

client = AsyncLLMKit(api_key="llmk_...")
completion, cost = await client.chat(
    model="gpt-4.1",
    messages=[{"role": "user", "content": "hello"}],
)

No SDK needed

LLMKit is OpenAI-compatible. Any client works:

from openai import OpenAI

client = OpenAI(
    base_url="https://llmkit-proxy.smigolsmigol.workers.dev/v1",
    api_key="llmk_...",
)

Part of LLMKit

This is the Python SDK for LLMKit, an open-source AI API gateway. The full platform includes:

• Proxy (Cloudflare Workers) - budget enforcement, cost tracking, provider routing
• Dashboard (Next.js) - analytics, API key management, budget configuration
• MCP server - 11 tools for Claude Code, Cursor, and Cline cost tracking
• TypeScript SDK - same features for Node.js/Deno/Bun
• CLI - wrap any command with llmkit wrap -- node agent.js

License

MIT