SentinelGate
SentinelGate drops prompt injections at Token 0 before LLM inference, sealing actions with tamper-evident evidence bags.
Last verified:
What is SentinelGate?
SentinelGate is an API proxy for LLM calls that sits between a client and an upstream model provider. Developers point an OpenAI-compatible SDK at its base URL and pass a sk-sent API key. The page says it screens prompts before upstream dispatch, halts requests it flags with an HTTP 403, and returns Ed25519-signed evidence records. It is aimed at teams running LLM workloads, including financial and payment use cases that need audit trails.
SentinelGate pricing
Pricing model: Freemium
Developer $19/month (50,000 requests/month); Pro $99/month (500,000 requests/month); Enterprise custom pricing
SentinelGate pros
- Drop-in OpenAI SDK compatibility: change base_url and API key, with no other code changes claimed
- Blocks flagged prompts at the proxy before upstream tokens are consumed, returning HTTP 403 kinematic_security_veto
- Signed attestation hashes and evidence bags for each transaction, with a stated mapping to ISO 20022 SplmtryData
- States a zero-retention design: prompts are not stored, and only hashes, signatures and billing counters are logged
SentinelGate cons
- The veto mechanism is described in proprietary 'thermodynamic kinematic' terms with no published method, benchmarks or independent evaluation
- The FAQ says requests are forwarded to GPT-4o, but the sample code uses a Llama model, so the supported upstream providers and models are unclear
- Developer and Pro tiers are priced publicly, but Enterprise pricing, deployment options and SLAs require a sales contact
- Some Enterprise features, such as the Linux eBPF kernel driver and bare-metal deployment, apply only to specific infrastructure
Frequently asked questions about SentinelGate
How is SentinelGate different from standard LLM gateways?
The page says standard gateways offer telemetry, caching and regex or secondary-LLM guardrails. SentinelGate says it evaluates prompts in memory against kinematic boundary states, stops execution at Token 0 when instability is detected, and signs each transaction with an Ed25519 attestation.
Does SentinelGate store or log prompt payloads?
According to the page, no. Prompts are evaluated in memory, forwarded to GPT-4o, and evicted when the socket closes. Only attestation hashes, Ed25519 signatures and billing counters are stored in PostgreSQL.
What latency does the veto add?
The page says the in-memory evaluation takes under a millisecond and total proxy routing latency is typically under 15 ms. These figures are the vendor's claims.
What happens when a prompt triggers a veto?
The proxy stops generation at Token 0, returns HTTP 403 kinematic_security_veto, and drops the socket before upstream tokens are consumed. The client receives an Ed25519-signed evidence bag with the state coordinates and attestation hash.
How is an API key provisioned after checkout?
After Stripe checkout, the user is redirected to a one-time reveal page that shows the secret key. The raw key is then purged, and only a salted SHA-256 hash is kept.
How do I rotate a key or get developer support?
Email [email protected]. The page says the team revokes the credential across clusters and issues a 24-hour one-time re-claim token.