Your AI agents reach tools, databases, and services through MCP servers. Strivio sits at that call boundary as your security gateway: it inspects every MCP tool call before it executes, enforces the policies you set, stops actions that policy marks as requiring approval, and writes every decision to a tamper-evident log.
No signup required for the demo · Model-agnostic · SDK and API
Instructs the model to discard the instructions it was given.
Why it was flagged: Instruction-override language is the primary way an agent's behaviour is hijacked.
Recommended action: BLOCK
The Model Context Protocol made it easy to give agents real tools: file systems, databases, payment APIs, internal services. Most of those connections were set up for speed, not control. The result is a familiar picture in production today:
Agents call MCP servers directly. Nothing inspects a tool call between the model's decision and the system's execution, so a manipulated prompt becomes a real action.
Tool access often runs on shared API keys and broad permissions. When an agent acts, the logs name a service account or a human's token, not the agent or the decision behind it.
After an incident, teams need three answers: what did the agent do, why did it do it, and was it allowed to? Scattered server logs rarely reconstruct that story.
Strivio integrates at your agent's call boundary through the SDK or API, so MCP tool calls pass through the same four steps as every other agent action.
Every prompt, tool call, and response is inspected in real time before it reaches its destination, including the arguments the agent wants to send.
The policy engine checks the call against your rules: what this agent may touch, what it may do, and what needs a person. Risk is scored on every interaction.
Allow, block, or flag for review, or stop a call that policy marks as approval-required. Prompt injection and data exfiltration attempts are blocked inline, not discovered later.
Every decision lands in a tamper-evident audit log: the agent, the call and the decision.
Policies are written around the actions your agents take, not around network traffic. Each rule names the actions it covers and the effect: allow, deny, or require approval. Separately, detected personal data such as email addresses and U.S. phone numbers can be redacted before a tool call is forwarded. Examples teams start with:
Payments and transfers are routine until they are not. Name the actions an agent must never take on its own.
actions: transfer_funds, make_payment effect: REQUIRE_APPROVAL
Some tools should never run from an agent session, no matter what the prompt says.
actions: drop_table, delete_records, destroy_infra effect: DENY
Reads are fine in the systems an agent serves. Sending customer data outside is a different decision.
actions: data_export dataClasses: PII_SSN, PII_EMAIL, PII_PHONE_US effect: DENY
Each gateway installation carries the list of MCP servers and tools it may use. A tool that is not on that list, or whose definition changed, is refused before any policy runs.
tool not in the approved list → refused and recorded
Integrate at the call boundary with the Strivio SDK or the API, and enforce on every tool call, including calls routed to MCP servers.
A callback adapter brings Strivio's inspection, enforcement, and logging to LangChain agents without restructuring your application.
Strivio governs the action layer, not the model. Keep the models and frameworks you have; the policies follow the agent, not the vendor.
New to MCP gateways? Start with the guide: What is an MCP gateway?
Paste a prompt injection attempt or a data exfiltration request into the live demo and watch it get scored and classified. No signup, no sales call required. Plans start at $499/mo when you are ready to enforce on your own agents.