Security and safety infrastructure for AI agents. MIT licensed. Pure Python.
When you run AI agents autonomously — trading bots, code assistants, document processors — three things break silently: costs explode from retry loops, inputs get poisoned by adversarial prompts, and decisions become untraceable when something fails at 3AM.
We built the fix. Five packages. All MIT. All offline-capable.
| Library | What it does | Install |
|---|---|---|
| ai-cost-guard | Hard budget cap — blocks LLM calls before they exceed your limit | pip install ai-cost-guard |
| ai-injection-guard | Prompt injection scanner — 22 patterns, 5 categories, entirely offline | pip install ai-injection-guard |
| ai-decision-tracer | Decision logger — JSONL auto-save, Markdown reports, built-in CLI | pip install ai-decision-tracer |
| agent-safety-mcp | MCP server — gives any AI assistant access to all three tools above | pip install agent-safety-mcp |
| mcp-security-audit | Security auditor — scan any MCP server for injection patterns, risky tools, misconfigs | pip install mcp-security-audit |
from ai_cost_guard import CostGuard
from prompt_shield import PromptScanner
from ai_trace import Tracer
guard = CostGuard(weekly_budget_usd=5.00)
scanner = PromptScanner(threshold="MEDIUM")
tracer = Tracer("my-agent", meta={"model": "claude-haiku-4-5"})
@guard.protect(model="anthropic/claude-haiku-4-5-20251001")
@scanner.protect(arg_name="prompt")
def call_llm(prompt):
with tracer.step("llm_call", prompt_len=len(prompt)) as step:
response = client.messages.create(...)
step.log(tokens=response.usage.input_tokens)
return responseBuilt from running autonomous AI trading agents in live financial markets. Every package exists because something broke in production.
