Voyager Browser

@dir-ai/voyager-browser

Voyager's web-page sense. A safe, read-only observation of one live URL: its structure, forms, links, script origins, security posture (HTTPS / HSTS / CSP / mixed-content / cookies) and accessibility signals (lang, image alt coverage, heading order) — each with a described, never applied fix.

It parses static HTML — no JavaScript execution, no headless browser — so it's honest about what it can and cannot see (it will not render a client-side SPA's runtime state). Every piece of page text is returned framed as untrusted. It is SSRF-gated: it refuses non-http(s) URLs and anything resolving to private, loopback, or cloud-metadata addresses.

One sense in the Voyager family. Read-only, like all the senses — it never submits a form, clicks, or mutates anything.

Install

npm i -g @dir-ai/voyager-browser

Use

voyager-browser observe https://example.comvoyager-browser observe https://example.com --json

https://example.com/https://example.com — 200; 3 finding(s), worst: medium. 0 form(s), 1 link(s).
  title: Example Domain  security: https  ·  a11y lang:yes alt:n/a
  med   HTTPS without HSTS        ↳ add Strict-Transport-Security: max-age=63072000; includeSubDomains  med   no Content-Security-Policy        ↳ add a Content-Security-Policy to constrain scripts/resources  low   no X-Content-Type-Options: nosniff        ↳ add X-Content-Type-Options: nosniff

As an MCP server

One tool, observe_page.

{ "command": "voyager-browser", "args": ["mcp"] }

As a library

import { observe } from '@dir-ai/voyager-browser'
const brief = await observe('https://example.com')console.log(brief.summary)console.log(brief.security)   // { https, hsts, csp, mixedContent, thirdPartyScripts, … }console.log(brief.forms)      // insecure / cross-origin / sensitive flagsconsole.log(brief.findings)   // severity + described fix

What it looks for

• Security posture (quality, not just presence) — plain HTTP; missing/weak HSTS; missing CSP or a CSP graded weak (unsafe-inline/unsafe-eval/wildcard/no object-src/no base-uri); missing clickjacking protection (X-Frame-Options / frame-ancestors); missing nosniff, Referrer-Policy, Permissions-Policy; version-leaking Server/X-Powered-By; mixed content on an HTTPS page (origin-compared, not prefix); per-cookie Secure/HttpOnly; third-party scripts without Subresource Integrity.
• Forms — a form on HTTPS posting to plain HTTP (critical); a sensitive form (password/payment) posting cross-origin or over HTTP; a sensitive POST with no anti-CSRF token; sensitive data on GET.
• Links — external target="_blank" without rel="noopener" (reverse-tabnabbing).
• Body-content leaks — directory listings (Apache/nginx autoindex), language-specific stack traces (Python/Java/PHP/Node/Oracle/SQLSTATE/MySQL) and verbose framework debug pages (Werkzeug/Flask, Rails, Symfony/Whoops, ASP.NET YSOD) disclosed in the response body — each with the matched signature (framed).
• Exposed JWTs — JWT-shaped tokens found in the HTML, response headers, Set-Cookie, and same-origin bundles are decoded (header+payload, base64url, no signature verification, no secret cracking) and flagged for alg:none (critical), expired (exp in the past), or missing exp. Claim values are framed as untrusted.
• Passive discovery of well-known sensitive paths — bounded, same-origin, read-only GETs to a short fixed list (/.git/config, /.env, /.svn/entries, /.DS_Store, /config.json, /wp-config.php~, /backup/, /uploads/), flagging only a confirmed body signature — never a bare 200. Pinned to the vetted IP; honours --authorized. Disable with --no-discovery.
• Accessibility — missing <html lang>, images without alt, skipped heading levels, unlabeled form fields.
• Render honesty — a render: static | hybrid | client-heavy field: if a page's content is JavaScript-rendered, the brief says so and marks itself PARTIAL rather than reporting a shell as clean.
• Composition hints — third-party script origins to vet with @dir-ai/voyager / @dir-ai/voyager-net.

Safety

• Read-only. Fetches the page with a single GET; never submits, clicks, or mutates.
• No code execution. Static HTML parsing only — the page's JavaScript is never run.
• SSRF-gated, with IP pinning. Only http(s); a single URL (no lists/credentials). The host is resolved, every address is classified canonically (IPv4-mapped IPv6, unspecified, CGNAT, NAT64, link-local, private, loopback, metadata all refused), and the connection is pinned to the vetted IP so DNS rebinding cannot swap in an internal address between the check and the fetch. Every redirect hop is re-vetted and re-pinned.
• Untrusted by construction. Every owner-controlled string that reaches the brief — title, headings, meta, links, form actions, field names — is injection-stripped and framed; the agent must treat it as data, not instructions.
• Bounded. One deadline covers headers and body; the body is byte-capped; a truncated read downgrades confidence and is never reported as "clean".
• Honest limits. It reads what the server sends. It does not see client-rendered state — and the render field says exactly how much it saw.

Roadmap

• v0.3 — same-origin JS-bundle static scan (exposed secrets/endpoints/source-maps), third-party/tracker/cookie inventory + tech fingerprinting, WCAG-mapped a11y depth, and a CognitiveClaim adapter so a page observation drops into a @dir-ai/voyager-agent mission (page → host → dependency chain).
• v1.0 — an opt-in, consent-gated --render sandboxed headless pass (network-isolated, resource-capped) for true SPA/rendered-DOM coverage, always labeled as render-mode output; a stable finding-kind taxonomy and a full posture score.

The line stays fixed: voyager-browser expands by reading more of what's already served, never by doing more to the server. Anything active (submitting, fuzzing, probing) belongs to a separate consent-gated organ.

The Voyager family

License

MIT © dir-ai