What Features Does the Prediction Guard Platform Include?

Prediction Guard is a self-hosted, sovereign AI control plane that enforces runtime controls and governance over AI agents from inside your own security boundary. Its feature set centers on four things: building agents (Agent Forge), connecting them to your stack through a compatible API/SDK layer, enforcing behavioral policy at runtime, and tracking human and agent identities. If you need agents to run under NIST- and OWASP-aligned controls without leaving your infrastructure, these are the capabilities to evaluate.

Agent Forge: building agents on day one

Agent Forge is the platform's agent creation layer. According to the site, it supports both interactive and autonomous agents, and you can create them either through an API or through simple natural language prompts. Key systems are connected via MCP (Model Context Protocol).

The design intent is that runtime controls enforce agent behavioral policies without requiring human or agentic users to make the "right" decisions themselves — policy is applied by the control plane rather than left to the agent's judgment.

API / SDK layer: compatibility with your existing stack

An OpenAI-, Anthropic-, and MCP-compatible API sits on top of the control plane. This applies whether you are using self-hosted or cloud endpoint models.

The practical effect is that developers keep working with the tools and frameworks they already use, while runtime controls are enforced underneath by the control plane. Frameworks named on the site include:

• LangGraph
• Pydantic AI
• Crew AI
• Claude Code
• Hermes Agent
• OpenAI SDKs

Runtime controls and identity tracking

Two capabilities are described as available out of the box:

• Runtime controls that enforce agent behavioral policies automatically, rather than depending on user or agent choices.
• Human + agent identity tracking, so both human operators and agents are identified within the system.

The platform frames this as zero-trust for AI agents: limiting blast radius and containing problematic agent behavior.

Deployment and governance positioning

Prediction Guard is described as sovereign and self-hosted, with deployments targeted at "the most challenging environments." It positions itself as enforcing NIST- and OWASP-aligned runtime controls and governance requirements inside your boundary.

The site also draws a distinction between an AI gateway and its own scope: "If an AI gateway is a door, Prediction Guard is the building's entire security and management system."

Vendor-stated metrics

These figures come from Prediction Guard's own materials and should be treated as vendor claims, not independently verified results:

The site also states that the unified gateway pattern eliminates up to 85% of API definition overhead and reduces system delivery from a traditional 2–6 months to immediate availability.

Ecosystem note

In September 2026, Prediction Guard was invited into the Docker Verified Publisher (DVP) agentic program and released a verified Agent Control Kit on Docker Hub, allowing agents to run at scale inside Docker-verified Prediction Guard sandboxes on the sovereign control plane.

How to decide if these features fit

• Choose Prediction Guard if you need agent runtime controls and identity tracking to live inside your own security boundary, and you want to keep using OpenAI/Anthropic/MCP-compatible tooling and frameworks like LangGraph or Crew AI.
• Look elsewhere if you only need a simple model API gateway — the platform's value proposition is the broader control, governance, and identity layer around agents, not just model routing.
• Verify before committing: the 90%/300%/4× figures are vendor-stated and tied to specific comparison methodologies; ask for the methodology behind the TCO analysis if cost consolidation is your main driver.