Head to head · updated 2026-07

AEGIS vs Prediction Guard

Both position as sovereign AI control planes. Both do runtime enforcement. The difference is architecture (Merkle audit + workflow anchoring + compensator rollback) and license posture (open-source vs commercial).

Written by the AEGIS team. Read Prediction Guard's docs and decide for yourself.

The 8 things buyers ask about first

Capability

AEGIS

Prediction Guard

Runtime intercept (block / pending)

Both gate tool + LLM calls before execution.

yes

Cryptographic audit chain (Merkle)

AEGIS ships Ed25519-signed transparency log. Prediction Guard logs policy decisions but not to a hash-chained tamper-evident structure.

yes

Workflow-anchored policies

binding_uuid > node_uuid > tool_name precedence in AEGIS. Prediction Guard scopes policies at the model / endpoint level, not the workflow node.

yes

Compensator + rollback

AEGIS ships a saga engine with binding-scoped compensators. PG stops at block; no undo primitive.

yes

EU AI Act evidence pack

Both position for EU AI Act. AEGIS ships article-by-article signed JSON export; PG offers policy alignment mapping but not a one-click evidence artifact.

yes

partial

Preset compliance DSL packs

AEGIS ships GLBA/PCI/SOX/DORA/HIPAA/FedRAMP as one-command imports with regulation citations. PG ships policy templates but fewer regulation-cited packs.

6 packs

partial

Air-gapped deploy

Both offer self-host and air-gapped. AEGIS additionally enforces at CNI layer via NetworkPolicy in the shipped Helm chart.

yes

Open source

AEGIS core is Apache/MIT. Prediction Guard is closed-source with a commercial license.

MIT (core)

closed

Where the two products diverge

Open source vs commercial license

AEGIS core is fully MIT/Apache — clone the repo, run the gateway, read every policy decision line yourself. Paid Cloud is optional. Prediction Guard is commercial-only; you pay for a license before you can run it, and you can't inspect the enforcement logic in production. Bespoke deployments in regulated environments increasingly need "we can read the source" as a hard checkbox.

Merkle-chained audit vs plain-log audit

Both write policy-decision logs. AEGIS additionally hashes each entry into a per-agent chain and Ed25519-signs the transparency-log root. If a bad actor with DB write access wants to rewrite history, they need the signing key too — Prediction Guard's approach records decisions but does not currently commit to a chain that would surface tampering. For SOC 2 Type II + auditor-signed-off evidence, the chain matters.

Workflow anchoring + rollback

AEGIS's rules bind to specific workflow node UUIDs (binding_uuid → node_uuid → tool_name precedence). Prediction Guard's policies attach at the model/endpoint level — same rule fires uniformly across every use of that endpoint. Same-tool-different-node differentiation isn't a primitive. AEGIS also ships a saga engine with binding-scoped compensators; PG stops at "block."

Full capability matrix

Capability

AEGIS

Prediction Guard

Runtime intercept (block / pending)

Both gate tool + LLM calls before execution.

yes

Cryptographic audit chain (Merkle)

AEGIS ships Ed25519-signed transparency log. Prediction Guard logs policy decisions but not to a hash-chained tamper-evident structure.

yes

Workflow-anchored policies

binding_uuid > node_uuid > tool_name precedence in AEGIS. Prediction Guard scopes policies at the model / endpoint level, not the workflow node.

yes

Compensator + rollback

AEGIS ships a saga engine with binding-scoped compensators. PG stops at block; no undo primitive.

yes

EU AI Act evidence pack

Both position for EU AI Act. AEGIS ships article-by-article signed JSON export; PG offers policy alignment mapping but not a one-click evidence artifact.

yes

partial

Preset compliance DSL packs

AEGIS ships GLBA/PCI/SOX/DORA/HIPAA/FedRAMP as one-command imports with regulation citations. PG ships policy templates but fewer regulation-cited packs.

6 packs

partial

Air-gapped deploy

Both offer self-host and air-gapped. AEGIS additionally enforces at CNI layer via NetworkPolicy in the shipped Helm chart.

yes

Open source

AEGIS core is Apache/MIT. Prediction Guard is closed-source with a commercial license.

MIT (core)

closed

Local LLM adapter (Ollama / vLLM)

Both support self-hosted models. AEGIS ships OpenAI-compat adapter for any /v1 endpoint.

yes

partial

NL-to-DSL compiler

AEGIS resolves to specific workflow node UUIDs; PG has an NL policy authoring path.

yes (workflow-aware)

yes

Cockpit / operator UI

Both have hosted admin UIs. AEGIS Cockpit is self-hostable open source.

shipped (Next.js)

shipped

Pricing surface

AEGIS free-forever core + paid Cloud. PG is commercial-only with per-seat annual license.

MIT + hosted

annual license

Honest picking guide

Pick Prediction Guard when

You want a fully managed control plane and are fine trading source visibility for turnkey ops.
Your buyer wants a single vendor SLA without a self-host escape hatch.
You don't need a Merkle-chained audit or workflow-anchored policies.

Pick AEGIS when

You need a signed, offline-verifiable evidence pack for regulated audits.
You want to read the source — either yourself or for the security team's review.
You need workflow-anchored policies or rollback, neither of which PG ships today.
You want free-forever core with paid Cloud as optional.

Try the AEGIS playground.

Compile an English rule and evaluate a tool call — no signup, all client-side.

Open playground → Pricing