Side-by-side scope comparison
| Dimension | Arcjet | ThumbGate |
|---|---|---|
| Install surface | Runtime SDK in your Node / Python / Deno / Bun web application | PreToolUse hook inside the developer's AI coding agent process |
| Traffic direction | Inbound — what reaches your application | Outbound — what the agent is about to do |
| What it blocks | Bots, rate-limit abuse, prompt-injection in user input, PII egress, WAF violations | rm -rf traversal, destructive SQL against non-test, git push --force, MCP tool calls to untrusted hosts, secret-carrying file writes |
| Framework coverage | Next.js, Express, Fastify, NestJS, Nuxt, Astro, React Router, Remix, SvelteKit, Bun, Deno, Python | Claude Code, Cursor, OpenAI Codex CLI, Google Gemini CLI, Sourcegraph Amp, Cline, OpenCode, Claude Desktop |
| Decision boundary | HTTP request middleware in your web server | PreToolUse hook in the agent runtime, before tool API fires |
| AI in the gate? | No (Arcjet ships deterministic rules + their Shield WAF; prompt-injection detection is pattern-based) | No (deterministic PreToolUse rule match + lesson DB; no model in the enforcement path) |
| Lesson promotion from feedback | No — rules are configured by the developer | Yes — thumbs-down on a bad tool call promotes to a prevention rule via Thompson Sampling |
| Best alongside | ThumbGate at the dev-agent layer | Arcjet at the application-inbound layer |
The shared architectural insight
Both products land on the same core decision: the gate runs deterministically, in your runtime, with no LLM in the enforcement path. Arcjet says it about their Shield WAF and rate-limit rules. ThumbGate says it about the PreToolUse hook. Neither product asks an external "judge model" to decide if an action is safe — both run pattern-match + policy logic in-process, which is what makes them auditable, cheap, and survivable under load.
The vendors who put an LLM in the enforcement path lose on three axes at once: latency (every request waits for a model call), cost (every request pays for inference), and auditability (the model's decision is non-deterministic, so an audit log of "the model said it was fine" is not a defense). Arcjet and ThumbGate independently arrived at the same posture from opposite ends of the perimeter.
The dual-deploy story for a regulated firm
Take a fintech or law firm running its own customer-facing application and developing it with AI coding agents:
- Arcjet on the customer-facing app. Bot detection on the signup endpoint, rate-limit on the chat endpoint, prompt-injection scoring on incoming user messages, PII detection on form submissions, WAF rules on every route.
- ThumbGate on the engineering team's AI coding agents. PreToolUse rules block destructive shell, enforce per-repo scope on the agent's tool calls, prevent privileged customer data from being sent to external LLMs during dev workflows, and turn each incident into a prevention rule the next sprint inherits automatically.
Neither layer overlaps with the other. Together they cover both the application's attack surface and the developer-agent's action surface — which is what our /ai-malpractice-prevention page describes for the legal-vertical case.
FAQ
Does Arcjet have a PreToolUse hook?
Not at the IDE-agent layer. Arcjet's "For Agents" surface (MCP server support, Arcjet Guards, Plugin, Skills, AI app protection) protects an application that hosts an AI agent — a chatbot endpoint, an MCP server, a tool-using API — from external misuse. ThumbGate runs upstream of that, inside the developer's coding agent before any tool call leaves the agent's memory.
Where does each one log evidence?
Arcjet emits decisions to your application's logging pipeline and the Arcjet dashboard for analytics. ThumbGate writes structured allow/warn/block decisions to a local lesson DB and (optionally on the Pro tier) syncs anonymized rule patterns to a hosted evidence dashboard. Both are SIEM-pluggable.
Can ThumbGate enforce policy on the application Arcjet protects?
No, and that is the point. ThumbGate runs in the dev's local agent runtime, not in the production web server. If an attacker hits your production app, Arcjet is the layer that sees the request first. If your AI coding agent is about to push to production, ThumbGate is the layer that sees the action first.
Pricing — what tier do I need from each?
Arcjet has a free tier and paid tiers for production volume (see arcjet.com/pricing). ThumbGate ships an open-source free tier with the full PreToolUse engine and prevention-rule promotion; Pro/Team adds hosted evidence sync, adapter coverage for all eight agent runtimes, and the audit-export endpoint we ship to procurement teams. The two pricing decisions are independent.
Is this comparison sponsored or partnered?
No. We don't have a partnership with Arcjet. We wrote this page because the same prospects evaluate both vendors — we want them to choose by scope, not by confusion. If anything here misrepresents Arcjet, open an issue at our repo and we will correct it.