AI Action & Spend Control

Every action your AI takes, checked before it happens.

Troxy sits between your AI agents and every action they attempt - sending emails, accessing data, running commands, making purchases. You set the policies, org-wide and per space. Troxy enforces them in milliseconds and tracks exactly what every agent costs, real billing data across providers, not just token math. Allow, block, escalate, or notify - before anything runs.

Free to start  ·  MCP-native  ·  Live today
terminal · getting started
# Install Troxy in your AI agent - takes 90 seconds npx troxy-cli init --key txy-... ✓ MCP configured. Restart your agent client. # Create policies in the dashboard, no code needed Agent wants to send emailALLOW (policy: business hours only) Agent wants to export 500 recordsESCALATE (policy: bulk data export) # Troxy intercepts every action your agent attempts Agent wants to delete production tableBLOCK (policy: destructive actions) Agent wants to create new userALLOW + NOTIFY (policy: admin actions) # Every decision logged. Approve from the dashboard in one tap.

The moment you give an AI agent the ability to act - send an email, query a database, run a command, make a purchase - you need a way to stay in control without micromanaging. Troxy enforces your rules quietly, holds anything that needs your eye, and lets you approve or deny from wherever you are.

One integration. Total control.
01
Agent triggers
AI agent attempts an action via MCP
02
Troxy intercepts
Policy engine evaluates in <100ms
03
Pass or escalate
Auto-approved, blocked, or sent to you
04
Audit trail
Every decision logged with full task context
~50ms
Average policy evaluation time. Your agents don't slow down.
4
Decision types: Allow, Block, Escalate, Notify. Everything else is a variation of these.
100%
Actions logged. Every decision has a full audit trail - agent, action, policy, timestamp.

Org & Space policies, org sets the ceiling

Org admins set the ceiling. Space admins can tighten it, never loosen it. No personal, self-serve policies, so nothing quietly slips through.

Real-time policy engine

Action filters, spend limits, daily activity caps, approval thresholds. Deterministic rules, sub-100ms decisions.

Human approval flows

Anything above your threshold gets held and sent to you. One tap to approve or decline. Agents wait, not you.

Prompt injection shield

Malicious prompts can't hijack your agent into unauthorized actions. Policy layer catches it before anything runs.

Full audit trail

Every decision logged with task context - what was attempted, why, by which agent, against which policy.

Real spend, every provider

Like a FinOps tool, but for what your agents actually cost. Real billing data from Anthropic and OpenAI, reconciled against token-math estimates and each tool's subscription plan, one true number instead of a guess.

Start checking every action today

No waitlist, no sales call. Create an account, connect your first agent, and set your first policy in a couple of minutes.

Free to start  ·  MCP-native  ·  Live today