Sentinel sits between your app and its model, inspecting every prompt before it becomes an instruction.
LLMs and agents treat everything they read as a potential instruction: web content, retrieved documents, tool results, not just what a user types in a chat box. Sentinel scrubs that content on the way in, and scans tool results on the way back to the agent, so the model only ever acts on what it should.
Grok · Gemini
What, who, when, and where
An AI Firewall, inline
Sentinel scrubs untrusted content — prompt injection, jailbreaks, data-exfiltration attempts — before your model sees it, and scans tool results before they return to an agent.
- HTML and hidden-content extraction, then encoding decode, normalization, fast-path regex, and semantic similarity
- PII and secret redaction, run independently of the threat score (Teams & Enterprise)
- Package-hallucination checks that flag a model recommending a name that isn’t in the registry
- Every request resolves to one of four outcomes: clean, flagged, neutralized, or blocked
Four audiences, one enforcement point
Same product. Each group leans on a different part of it.
- Developers — one base URL swap, or one REST call
- Security / InfoSec teams — a single audit trail for AI usage org-wide
- Insurers / risk assessors — every request scored, categorized, and logged
- Compliance-driven orgs — PII and secrets redacted before content reaches a cloud model
Any time the model reads content it didn’t write
That includes, but isn’t limited to:
- Web fetches and crawled content
- Retrieved chunks from a RAG knowledge base
- End-user input in a chat or support flow
- Tool outputs, MCP server responses, and third-party agent skills
It matters most for agentic sessions that take real actions, and for any regulated data heading toward a cloud model.
Inline, self-hosted or SaaS
Nothing reaches the model unscored.
- Self-hosted — a Docker Compose stack on your own infrastructure. Content never leaves your network.
- SaaS — hosted at sentinelaifirewall.com, nothing to run yourself
- Integration is a REST call to /v1/scrub, an SDK base-URL swap, or the Clawhub sentinel-proxy skill
Built for how each team already works
Briefs & one-pagers
Each opens as a printable page. Use your browser's Save as PDF (or the button in the top corner of the page) to download a copy.
Product Brief
How the detection pipeline works, the current adversarial benchmark result (detection rate and false-positive rate), a per-category breakdown, and an honest note on the one attack class still on the roadmap.
Reseller Partner Brief
The market case for prompt-path security, what Sentinel covers, and how the white-label reseller program works — dedicated single-tenant stacks, co-branding options, and integration.
Overview one-pager
The short version of this page as a single printable sheet — the pipeline, the four outcomes, and the four audiences at a glance.
Start free, no card required
100 requests a month on Starter, forever free. Add a key, swap a base URL, and Sentinel is inline.