Everything you need to know about securing your AI infrastructure.
Do I need to rewrite my application code?
expand_more
No. Agent-M acts as a universal drop-in replacement for the official SDKs. You simply change your `baseURL` to point to `https://app.agent-m.ai/v1` and swap your provider API key for your secure Agent-M Proxy Key. Your existing payload remains completely untouched. We even natively translate OpenAI-formatted payloads to Anthropic or Gemini on the backend automatically.
Does Agent-M support open-source models and custom endpoints?
expand_more
Yes. Beyond native support for OpenAI, Anthropic, Gemini, Groq, and Mistral, you can configure an Agent-M proxy key to point to any custom OpenAI-compatible endpoint. This allows you to route traffic to your own local vLLM instances or private cloud clusters while maintaining all of our pre-execution firewall and telemetry features.
Will routing through a proxy break MCP or Agentic tool calling?
expand_more
No. Unlike legacy proxies that use heavy, proprietary SDKs which often strip metadata, Agent-M is built specifically for the Agentic era. We natively pass Model Context Protocol (MCP) tool calls, system prompts, and complex JSON schemas perfectly intact without dropping a single frame of context.
How exactly do the budget firewalls work?
expand_more
Most observability tools only alert you after a runaway agent has blown your budget. Agent-M is a Pre-Execution Firewall. When a request hits our edge, we calculate the payload size and check your exact remaining budget in milliseconds. If the request exceeds your set limit (e.g., $10/day), we physically sever the connection and return a 429 error before your provider is ever called, saving your wallet.
How does Smart Fallback Routing work?
expand_more
If you enable a Fallback Model on a proxy key, Agent-M watches the connection to the primary AI provider. If the provider goes down, returns a 529 overload, or hits a rate limit, Agent-M intercepts the failure and instantly re-routes the exact same prompt to your chosen fallback model (e.g., from Claude 3.5 Sonnet to GPT-4o). Your end-user app never even knows an error occurred.
Is my data secure? Do you read my prompts?
expand_more
We act as a blind proxy. We do not store, log, or train on your prompts, system messages, or AI completions. We only log the metadata required for your dashboard (timestamp, model used, latency, and token counts). Furthermore, your actual Provider API keys are encrypted at rest using AES-256 and are never exposed to the client-side application.
What happens when my 14-Day Free Trial ends?
expand_more
If you don't add billing info by the end of your trial, your account gracefully downgrades to the free Developer tier. Your primary proxy key will remain active with a 100,000 token/month routing limit. Additional proxy keys and advanced features (like IP Whitelisting and Fallback routing) will be temporarily paused until you upgrade.