Most "AI in a product" is a documentation bot with a chat skin, or a demo that promises to run your whole operation and quietly can't. Iris is neither. It's a focused, honest tool: it drafts your edge rules from plain English, answers questions about your account by reading the live API, searches your traffic logs from a plain-English query, explains a single request, and answers how-to questions from the docs — with every write gated behind a confirmation card and a blast-radius dry-run. This post is a straight account of what Iris actually does today, and just as importantly, what it does not. For the broader picture, start with the Iris hub.
What Iris does: turn English into reviewable rules
The core of Iris is rule generation. Describe the behavior you want and it drafts the corresponding rule for you to review before anything is enforced:
- Firewall rules — "block scanners hitting /admin" becomes a rule with the right match and a block action.
- Rate-limit rules — "5 login attempts per minute per IP" becomes a scoped rate limit with the right window and scope.
- Traffic-policy rules — WAF, JA3 bot management, CORS, JWT validation, redirects, header rewrites, caching.
- Routing rules — picks the right agent from your connected fleet for a match condition.
- Error-page templates — full HTML/CSS with the error tokens filled in.
- Metric alerts — "alert me if p95 latency goes over 300ms" becomes a threshold alert you enable.
Each of these is also an HTTP endpoint, so you can script rule drafts into onboarding or CI. The point is the same in the dashboard or over the API: Iris drafts, you review, and only then does it apply.
Log search, account questions, and per-request explain
The reading side is where Iris earns its keep, and it's precise rather than magical. Ask a plain-English question about your HTTP traffic logs — "blocked POSTs to /api from outside the EU in the last hour" — and Iris extracts the structured search parameters (method, status, path, IP, time range) and runs the search, handing you a real Traffic Inspector filter you can see, tweak, and re-run. It's a search accelerator grounded in your captured traffic, not an opaque answer.
Iris also answers questions about your account by reading the live API rather than guessing: "what endpoints do I have?", "who are my members and what are their roles?", "how close am I to my plan limit?", "is my account verified?". It calls the real endpoints, members, entitlement, and verification reads under your own permissions, so the answer reflects your account as it is right now — and it can compile a structured, downloadable account report (overview plus per-area tables) that you export as HTML or PDF client-side.
Per-request explain works one request at a time. Open a captured request in the Traffic Inspector and Iris tells you which rule matched and why it was blocked or allowed, classifies the client (bot, scanner, legitimate, misconfigured), and suggests a concrete fix. It reasons about that single request — it is not a one-click "explain this spike" narrative over your whole account.
Every write is gated
Iris can create, update, disable, delete, and reorder rules — but never silently. Every write surfaces a confirmation card with a field-level before-and-after diff, and firewall, rate-limit, and traffic-policy changes additionally get a blast-radius dry-run that replays the proposed change against your real recent traffic: how many requests it would newly block or throttle, how many were succeeding, and a safe / risky verdict. Nothing applies until you approve, and every AI tool call lands in the AI-activity ledger, so "what did the AI do?" always has an answer.
There is one opt-in exception, and it's deliberately narrow. Firewall-rule autonomy can be switched on per account so Iris applies a safe, reversible firewall change on its own after a passing dry-run — and auto-reverts it if real traffic regresses. It's firewall-only. Every other kind of write always comes back to you.
What Iris does not do (on purpose)
Being honest about the edges is part of trusting a tool that touches your config:
- Everything it reads is scoped to your permissions. Account questions and log searches run under your own role — Iris can't read past what you're allowed to see, and it reads the live API rather than guessing, so it won't invent an endpoint, member, or quota that isn't there.
- It does not diagnose outages in one call. There's no "why is api-prod down?" button. Per-request explain is scoped to one request you point it at.
- Autonomy is firewall-only. Reads are free, but writes beyond firewall rules always come back to you for approval. Iris can't widen its own permissions, and admin, internal, and irreversible account actions (deleting or transferring the account) are denied forever, prompt-independent.
Guest vs authenticated
In a guest chat — before you sign in — Iris knows the platform: how endpoints work, what a traffic policy does, how to attach a WAF, where to click. It's the knowledgeable-stranger mode. In an authenticated chat, the same assistant can draft rules on your account (behind the confirmation gate), answer questions about your endpoints, members, entitlements, and verification from the live API, search your traffic logs from a plain-English query, and explain your captured requests. Same model, same platform knowledge, plus the ability to read and act — always within your own role and always gated on writes.
Getting there costs nothing to try. The free plan gives you 5 endpoints with no credit card. Point a tunnel at a local service and start asking Iris to draft you a rule:
# install and expose a local app in one command
brew install ngris
ngris http 3000
# or bind it to a hostname you own
ngris http 3000 --url app.ngris.comOnce you have an endpoint, open the chat and ask for something real — "block scanners on /admin" — and watch it draft the rule, dry-run it, and wait for your approval. When you're ready to go deeper, generating a firewall policy with AI is the natural next step.