MCP server for AI agent safety, cost guards, and prompt injection scanning.
agent-safety-mcp
MCP server for AI agent safety. One install gives any MCP-compatible AI assistant access to cost guards, prompt injection scanning, and decision tracing.
Works with Claude Code, Cursor, Windsurf, Zed, and any MCP client.
Install
Claude Code (recommended)
claude mcp add agent-safety -- uvx agent-safety-mcp
Manual (any MCP client)
Add to your MCP config:
{
"mcpServers": {
"agent-safety": {
"command": "uvx",
"args": ["agent-safety-mcp"]
}
}
}
From PyPI
pip install agent-safety-mcp
agent-safety-mcp # runs stdio server
Tools
Cost Guard — Budget enforcement for LLM calls
| Tool | What it does |
|---|---|
cost_guard_configure |
Set weekly budget, alert threshold, dry-run mode |
cost_guard_status |
Check current spend vs budget |
cost_guard_check |
Pre-check if a model call is within budget |
cost_guard_record |
Record a completed call's token usage |
cost_guard_models |
List supported models with pricing |
Example: "Check if I can afford a GPT-4o call with 2000 input tokens"
Injection Guard — Prompt injection scanner
| Tool | What it does |
|---|---|
injection_scan |
Scan text for injection patterns (non-blocking) |
injection_check |
Scan + block if injection detected |
injection_patterns |
List all 75 built-in detection patterns across 9 categories |
Example: "Scan this user input for prompt injection: 'ignore previous instructions and...'"
Decision Tracer — Agent decision logging
| Tool | What it does |
|---|---|
trace_start |
Start a new trace session |
trace_step |
Log a decision step with context |
trace_summary |
Get session summary (steps, errors, timing) |
trace_save |
Save trace to JSON + Markdown files |
Example: "Start a trace for my analysis agent, then log each decision step"
What this wraps
This MCP server wraps the AI Agent Infrastructure Stack — three standalone Python libraries:
- ai-cost-guard —
pip install ai-cost-guard - ai-injection-guard —
pip install ai-injection-guard - ai-decision-tracer —
pip install ai-decision-tracer
All three: MIT licensed, zero runtime dependencies (individually), pure Python stdlib.
The MCP server adds mcp>=1.0.0 as a dependency for the protocol layer.
Why
AI coding assistants (Claude Code, Cursor, etc.) can now protect the agents they help build — checking budgets, scanning inputs, and tracing decisions — without leaving the IDE.
Built from 8 months of running autonomous AI trading agents in live financial markets.
License
MIT
Tools (12)
cost_guard_configureSet weekly budget, alert threshold, and dry-run mode.cost_guard_statusCheck current spend vs budget.cost_guard_checkPre-check if a model call is within budget.cost_guard_recordRecord a completed call's token usage.cost_guard_modelsList supported models with pricing.injection_scanScan text for injection patterns (non-blocking).injection_checkScan and block if injection detected.injection_patternsList all 75 built-in detection patterns across 9 categories.trace_startStart a new trace session.trace_stepLog a decision step with context.trace_summaryGet session summary including steps, errors, and timing.trace_saveSave trace to JSON and Markdown files.Configuration
{"mcpServers": {"agent-safety": {"command": "uvx", "args": ["agent-safety-mcp"]}}}