The runtime safety net for AI agents.

Budget limits, kill switches, and approval gates — inside your agent, across every provider.

---

Why Wickd?

Zero latency	Runs in-process. No proxy, no sidecar, no network hop.
43 models tracked	Real-time pricing for OpenAI, Anthropic, and Google models.
1 decorator	Add budget enforcement to any agent in one line of code.
Streaming-safe	Tracks cost from streaming responses, including Anthropic messages.stream().
Concurrent-safe	Per-run budget isolation via ContextVar / AsyncLocalStorage.
292 tests	Unit, integration, and real API tests across all 3 providers.

---

Quick Start

Python

pip install wickd-ai

import wickd
import openai

@wickd.agent(budget=wickd.Budget(per_run=0.50, daily=5.00))
def my_agent(task: str):
    client = openai.OpenAI()
    return client.chat.completions.create(
        model="gpt-4o",
        messages=[{"role": "user", "content": task}],
    )

try:
    result = my_agent.run("summarize this document")
except wickd.BudgetExceeded as e:
    print(f"Agent stopped: {e}")

TypeScript

npm install wickd

import { agent, Budget, BudgetExceeded } from "wickd";
import OpenAI from "openai";

const myAgent = agent({
  name: "my_agent",
  budget: new Budget({ perRun: 0.50, daily: 5.00 }),
  fn: async (task: string) => {
    const client = new OpenAI();
    return await client.chat.completions.create({
      model: "gpt-4o",
      messages: [{ role: "user", content: task }],
    });
  },
});

await myAgent.run("summarize this document");

Proxy Mode (zero code changes)

pip install wickd-proxy
wickd-proxy start --budget-per-run 0.50 --budget-daily 5.00

# Just set an env var — your existing code works unchanged
export OPENAI_BASE_URL=http://localhost:4319/openai/v1

---

Supported Providers

Provider	Models	Streaming	Async
OpenAI	GPT-4o, GPT-4.1, o1, o3, o4-mini, ...	✅	✅
Anthropic	Claude Opus 4, Sonnet 4, Haiku 4.5	✅	✅
Google	Gemini 2.5 Pro, 2.5 Flash	✅	✅

43 models with real-time pricing. Unknown models use conservative fallback estimates.

---

Features

Feature	Description
Budget enforcement	Per-run, daily, and monthly cost caps. Checked before the response reaches your code.
Kill switches	Automatic halt when spend exceeds limits. Raises BudgetExceeded immediately.
Approval gates	Pause execution for human review. Slack, webhook, terminal, or custom handlers.
Streaming support	Tracks cost from streaming responses. Auto-injects stream_options for OpenAI.
Tool tracking	Trace MCP tool calls alongside LLM requests. Approval gates on dangerous tools.
Patch verification	Runtime health checks confirm interception is active. Configurable: block, warn, or allow.
Transport fallback	Falls back to httpx-level interception when SDK patching fails.
Async agents	Full async/await support with arun(). Context isolation across concurrent runs.

---

How It Works

Wickd intercepts LLM SDK calls at the method level. No separate server. No network hop.

Your agent code
      │
      ▼
  Wickd interceptor (budget check → trace → cost)
      │
      ▼
  OpenAI / Anthropic / Google SDK
      │
      ▼
  LLM API

When your agent calls client.chat.completions.create(), Wickd's wrapper:

1. Checks budget before the call
2. Forwards to the real SDK
3. Reads token usage from the response
4. Calculates cost and records it
5. Checks budget again — kills the agent if exceeded
6. Returns the response to your code

---

Works With

Framework	Status
OpenAI SDK	✅
Anthropic SDK	✅
Google GenAI	✅
LangGraph	✅
CrewAI	✅
Vercel AI SDK	✅
OpenAI Agents SDK	✅
Any Python/TS code	✅

Wickd is framework-agnostic. It patches at the LLM SDK level, so it works with any agent framework or custom code.

---

Packages

Package	Install	Description
wickd	npm install wickd	TypeScript SDK
wickd-ai	pip install wickd-ai	Python SDK
wickd-core	npm install wickd-core	Shared types & pricing
wickd-proxy	pip install wickd-proxy	LLM proxy server

---

Contributing

Contributions welcome. Please open an issue first to discuss what you'd like to change.

git clone https://github.com/lundberga/wickd.git
cd wickd
npm install && npm run build          # TypeScript
cd packages/sdk-python && pip install -e ".[dev]"  # Python

Run tests:

npm run build && cd packages/sdk-typescript && npx vitest run   # TypeScript
cd packages/sdk-python && PYTHONPATH=. python -m pytest tests/  # Python
cd packages/proxy && PYTHONPATH=../sdk-python:. python -m pytest tests/  # Proxy

---

License

MIT