What is the difference between MAQPNA and Prediction Guard?
Prediction Guard is a commercial, self-hosted control plane that combines a governed, OpenAI-compatible model API with runtime controls for agents, such as PII handling, prompt-injection detection, MCP scoping, spend limits and audit logs. MAQPNA is a self-hosted agent runtime on Kubernetes: it runs each agent session in its own sandbox and puts one gateway in front of every tool, model, agent-to-agent and egress call, with default-deny policy, human approval with four-eyes quorum, DLP, a kill switch and a hash-chained audit ledger. If you want a governed model platform with injection and PII filters, Prediction Guard is the closer fit. If you need to decide, hold and prove individual agent actions, with the agent's code isolated and unable to bypass the gateway, MAQPNA is.
How do they compare side by side?
As of 1 November 2026, per each vendor's public docs. "Not described" means we found no description on Prediction Guard's public pages, not that the feature does not exist; ask them.
| Prediction Guard | MAQPNA | |
|---|---|---|
| Licence | Commercial. Self-serve billing through the AWS and Azure marketplaces, or an annual enterprise agreement (pricing) | Proprietary, open core. The Community edition is free for one cluster, in production too; paid editions are quoted (pricing) |
| Deployment model | A control plane that "runs inside your infrastructure beside your agents": self-hosted, in a cloud VPC or, on the Enterprise plan, air-gapped and single-tenant (home, pricing) | An operator, gateway and identity broker on your Kubernetes cluster, installed with Helm; maqpna dev on a laptop (install) |
| What it enforces | Policy checks on model interactions, tool calls and data access; PII block, log or mask; jailbreak and prompt-injection detection; per-MCP-server scope; token and spend limits (runtime governance) | Default-deny policy on every MCP tool call, model call, A2A call and egress connection, with rules on tools, arguments, users, groups, schedules and session taint; DLP; budgets; tool pinning (policies) |
| Where code runs | Your agents run where you run them; the control plane governs their calls | One sandbox per session (gVisor, Kata with Firecracker microVMs or confidential VMs) with default-deny egress except to the gateway (network and egress) |
| Human approvals | Not described as an approval workflow on its public pages | A rule can hold one call for a person: four-eyes quorum, no self-approval, user approval on their own device (CIBA), decisions in the CLI, console, Slack or Teams (approvals) |
| Audit | "Append-only, tamper-resistant storage" for control decisions, logs tied to agent identity, evidence bundles, streaming to Grafana, Splunk or Datadog (agent operations) | A SHA-256 hash-chained ledger with signed checkpoints, verifiable offline; SIEM and write-once storage connectors in the Enterprise edition (audit ledger) |
| Stopping agents | "Model kill switches" pushed from a central console (supply chain) | A kill switch that revokes an agent, session, user, token, namespace or everything across gateway replicas (kill switch) |
| Model hosting | Yes: an "OpenAI, Anthropic, and MCP compatible API" over self-hosted or cloud models | No model hosting; governs calls to any OpenAI-compatible model route. Local models are available through the separate llamay component of the Kosa8 Labs Stack |
| Air-gapped | Yes, on the Enterprise plan, per their pricing page | Yes: a signed offline bundle, offline licence verification, no phone-home (air-gapped install) |
| Maintained by | Prediction Guard | Kosa8 Labs |
What is Prediction Guard?
Prediction Guard describes itself as a "lightweight control plane" that gives regulated enterprises operational control of their AI agents and runs "inside your infrastructure beside your agents" (predictionguard.com). Its public pages, as of 1 November 2026, describe four areas:
- Runtime governance. Model interactions, tool calls and data access are checked before they run, with PII controls (block, log, mask or replace), detection of jailbreaks and direct and indirect prompt injection, and per-MCP-server scoping such as read-only database queries (runtime governance).
- Supply chain. An inventory of the models, MCP servers and tools each agent can reach, a built-in tool proxy with per-user OAuth 2.1 for MCP servers, per-system API keys and quotas, and an AI bill of materials in CycloneDX format (supply chain).
- Agent operations. Audit logs tied to agent identity, evidence bundles and real-time streaming of security events (agent operations).
- A model API and agent builder. An OpenAI-, Anthropic- and MCP-compatible API, with guardrail endpoints for PII, injection and toxicity in the developer docs, and an agent builder called Agent Forge.
What is MAQPNA?
MAQPNA is the sovereign runtime for AI agents. Each agent session runs in its own sandbox, gets a short-lived signed identity that names the agent and the person it acts for, and can only reach the outside world through the MAQPNA gateway. The gateway checks every MCP tool call, model call, agent-to-agent call and browser egress connection against policy, and records each decision in a tamper-evident ledger. It does not host models or build agents: your agents keep their framework and point their MCP and model clients at the gateway. See what is MAQPNA and what an MCP gateway does.
When should you choose Prediction Guard instead?
- You want models and governance in one product. Prediction Guard serves a compatible model API over self-hosted or cloud models. MAQPNA governs model calls but does not serve models.
- You want content filters on prompts and outputs. Prompt-injection and jailbreak detection and toxicity checks are part of its documented feature set. MAQPNA's in-process DLP is deterministic pattern detection for secrets and personal data; it is not a toxicity or jailbreak classifier.
- You do not run Kubernetes. MAQPNA's production install is a Kubernetes operator and Helm chart. If you have no cluster and do not want one, that is a real cost.
- You buy through cloud marketplaces. Prediction Guard's self-serve plan is billed through the AWS and Azure marketplaces, which can be the quickest procurement route.
- You want an agent builder. Agent Forge builds agents through an API or natural-language prompts. MAQPNA does not build agents.
When does MAQPNA fit?
- Agents run code or browse. MAQPNA isolates each session in a gVisor sandbox, a microVM or a confidential VM, with default-deny egress. See containers vs sandboxes.
- Some actions need a person. A rule can hold one tool call, bound to its exact arguments, for four-eyes approval with no self-approval. See human-in-the-loop approval.
- You must prove what happened. The ledger is a hash chain with signed checkpoints that a third party can verify offline. See tamper-evident audit trails.
- Policy must reach arguments and history. Rules match argument values, the person's groups and session taint, so "no external email after reading an untrusted web page" is one rule. See prompt injection in tool calls.
- You want to start free. The Community edition is the full runtime for one cluster at no cost.
The two are not mutually exclusive. An agent can call a model through any OpenAI-compatible endpoint while its tool calls pass through the MAQPNA gateway. We have not tested a combined deployment, so treat that as an architecture option, not an integration.