AginxBrowser

English | 中���

The Browser for AI Agents. See the live web. Read it. Act on it. Remember it.

English | 中文文档

A browser built for agents from the first line of code — not a human browser bolted onto automation. See the world, read it, search it, act on it, and keep what you read: one Rust binary with built-in V8, no Chromium required.

Humans have Chrome. Agents have AginxBrowser.

One binary, zero dependencies, instant service. HTTP API + native MCP + CDP — agents plug in and go, and existing Playwright / Puppeteer / browser-use code attaches directly.

Real pages rendered by AginxBrowser's diting engine (no Chromium) — Wikipedia, this repo, Rust. Screenshot it yourself →

Why Agents Need Their Own Browser

Measured against headless Chrome on the same 20 pages, same network (bench, 2026-08-28): 7.6× faster to agent-usable text (p50 532 ms vs 4 053 ms), ~10× less memory (227 MB for the whole process vs ~2.1 GB per Chrome page), and 0 hard failures where Chrome's --dump-dom produced no DOM on 5 of 40 loads. An agent's total cost is browser efficiency × model efficiency — this is the browser half.

Existing "browser automation" was built for humans or for one-shot scraping — not for agents:

An agent needs five things from a browser: see, read, find, act, remember. One binary covers them all — systemd-friendly, MCP-native for Claude/Cursor, zero dependencies.

Core advantage: no Chromium. AginxBrowser inlines a full browser engine (V8 + Rust HTTP stack + the diting CSS/layout/paint rendering engine, with the Blitz/Stylo/Taffy lineage as its reference implementation). No Puppeteer, no Chrome, no Docker. One Rust binary under systemd is your agent browsing infrastructure.

Three Things Stateless Renderers Can't Do

Most new "agent browsers" are stateless, fingerprint-less one-shot renderers — fine for public pages, dead on arrival against Cloudflare or login flows. AginxBrowser goes the opposite way:

• 🔐 Real TLS fingerprints — stealth mode replicates the complete Chrome145 / Firefox133 / Safari / Edge TLS handshakes via BoringSSL (not just a UA string), switchable per request; Cloudflare Turnstile challenges wait automatically for cf_clearance. Fingerprint-less engines eat 403s — we get through.
• 🤝 Stateful interactive sessions — login state injectable and exportable (session_create(cookies=...) ↔ session_cookies), surviving pagination and multi-step flows; persistent: true even survives idle eviction and server restarts — the same session id comes back logged in. One-shot engines throw state away.
• 🔌 MCP native — 37 tools as first-class citizens (not a CDP shim). Claude Code / Cursor / Claude Desktop connect in one line. HTTP + MCP dual protocol — plus a CDP bridge, so the DevTools ecosystem works too.

Reference point: Cloudflare's Kitesurf explicitly ships neither real TLS-fingerprint negotiation nor persistent auth sessions — anti-bot and login territory is exactly where AginxBrowser plays.

Apache-2.0 open source, single binary — self-host today, no cloud lock-in.

Every Fetch Is a Receipt

Agents act on what a browser tells them, so the response reports what actually happened — not just "got a 200":

• tier — which path served the page: plain HTTP (~100 ms) or the V8-rendered browser tier. An agent can see why a fetch was fast or slow.
• redirected_from — the full redirect trail. redirected_from[0] is the URL you asked for, url is where the content actually came from — requested paired with effective, every hop visible.
• content_hash + changed_since_prev — every fetch is hashed; consecutive samples of the same URL can be diffed. A rate-limited origin serving the same frozen 200 body for days reads as changed_since_prev: false — the cheapest drift detector there is.
• captcha_event — when a challenge page was detected (and solved, if a solver is configured), the response says so instead of handing over a challenge page as if it were content.

The local cache builds on the same idea: search hits come back with [§ heading] section prefixes so an agent knows where on the page a hit landed, and ranking fuses keyword relevance with freshness.

Capabilities

• Tiered rendering: static pages over plain HTTP (~100ms); V8 spins up only when JS rendering is needed (~1-2s) — 90% of the bench page set served without spinning up V8 at all; every response reports which tier served it (tier field)
• Multi-engine meta-search: general web (Baidu / Bing / Sogou / WeChat / Google / DuckDuckGo), news (Bing News), code (Stack Overflow, GitHub), packages (npm, PyPI), academic (arXiv), AI models (Hugging Face) — 15 engines across 7 categories, queried concurrently, merged and deduplicated. Operators can plug a private Meilisearch index into the same /search. Search → read in one step
• Image search: categories=images hits Baidu/Bing image indexes and returns direct binary image_url links (downloadable straight to jpg/png) plus source_url provenance
• Interactive sessions: persistent browser sessions with indexed interaction (state/click/input/scroll/eval) — agents browse like humans do, and session_export turns what an agent figured out into a runnable curl replay script (zero model tokens on re-run) — or, with format=json, into a flow document (flow_run replays it server-side with {{var}} substitution, wait/expect gates and saved outputs; installed flows live in workflow/<name>/flow.json, dropped in without a rebuild). Session tools also cover the acting part: session_viewport simulates device viewports (media queries respond), session_wait blocks on a selector or predicate with a timeout, session_screenshot renders the live state, session_console replays the page's console ring, and session_storage exports/restores cookies plus localStorage for login hand-off
• Playback-link sniffer: session_network(filter=media) extracts the m3u8/mp4/dash URLs a page's player actually requested at runtime — links found only in page HTML are often decoys, so the request log is the source of truth. GET /session/{id}/har exports the same traffic as HAR 1.2 (retained bodies included)
• CDP bridge: /json/version + /devtools/{kind}/{id} WebSocket — chromium.connectOverCDP() from Playwright, Puppeteer, or browser-use attaches with one line, and agent-browser drives it via --cdp (snapshot returns the synthesized accessibility tree with @ref handles) (integration guide). DevTools ecosystem compatibility without becoming a CDP shim
• File download: streaming to disk (no memory buffering), SHA-256 integrity, resume of interrupted transfers — for binaries, archives, datasets
• Local cache that remembers: every fetch/search lands in SQLite (FTS5) at ~/.aginxbrowser/cache.db. The cache tool re-answers from what the agent already read instead of re-paying network time: full-text search with CJK substring matching, keyword × freshness fusion ranking, [§ heading] section-aware snippets, per-URL content hashes for drift detection, TTL-bounded, per-session scoping for shared deployments
• CAPTCHA handling: type detection with automatic Cloudflare challenge wait and optional 2captcha integration — search never stalls on verification pages
• JS data extraction: js_extract pulls window.__INITIAL_STATE__ and other structured data out of SPAs
• Document generation: render_markdown turns markdown into a deterministic, self-contained HTML artifact — the document layer, so agents never write HTML by hand. Prose rides a plain offline shell (no fonts, no scripts); fenced archify blocks carry typed zero-coordinate diagram JSON (sequence / workflow / architecture / dataflow / lifecycle families) and render to inline SVG via the layout engine. Same input, same bytes — the receipt carries the sha256 so determinism is verifiable. theme (light/dark) and preset (classic / signal-flow / blueprint / editorial) bake colors at generation time; quality: "showcase" is the delivery gate, grading route crossings, label clearance and rhythm without touching the artifact bytes. Guided-view tabs plus window.agxViewer (focus / ego / route / reach) make the artifact interactive; with session_id it loads into a live session and the reply grades how it fits the viewport (fits/tall/wide/oversized). Mermaid sources are the agent's job to translate into archify JSON, not the engine's. Diagram vocabulary adapted from archify (MIT)
• Screenshot rendering: /screenshot endpoint (opt-in --features screenshot) paints the JS-rendered DOM with the diting rendering engine — pure CPU, no Chromium — to PNG. Vision input for agents
• Timeline video: /video endpoint + render_video MCP tool render a page's animation timelines to MP4 — the page's scripts register GSAP-style timelines in window.__timelines (duration() + pause(t)), each frame seeks to t=i/fps and paints the viewport, and the frames pipe into ffmpeg (H.264, yuv420p). Deterministic by construction: no wall clock in the pixel values, same render twice = same MP4. Needs ffmpeg on PATH
• Page set (PDF/PNG/PPTX/DOCX): /pdf endpoint + render_pdf MCP tool cut a rendered page into pages and package them — print mode paginates at top-level block boundaries (default A4 @96dpi, no half-cut text where a break can land on a block edge), slides mode makes one page per CSS-selector match sized to the element (an HTML deck with one .slide per page exports as a real deck). Image-based PDF: per-page JPEG via DCTDecode, hand-rolled PDF 1.4 writer, zero new dependencies. PPTX packages the same pages as one slide per page; DOCX as one page-sized section per page — both hand-rolled OOXML (stored-ZIP writer, fixed timestamps), byte-deterministic, zero new dependencies
• TLS fingerprint spoofing: stealth mode impersonates Chrome145/Firefox133/Safari/Edge, switchable per request
• MCP server: --mcp mode exposes 37 tools (fetch/eval/search/download/cache + session + flow + screenshot + video/pdf + docgen tools) — Claude Code / Claude Desktop / Cursor call them directly
• Firecrawl compatible: /v1/scrape endpoint — existing Firecrawl clients migrate by changing the base URL
• DNS rebinding protection: built-in SSRF guard + post-resolution IP validation

A Browser, Not a Crawler

AginxBrowser exists for real-time retrieval: an agent arrives with a question, reads a handful of pages, leaves with the answer. It is not a crawling tool — and the product is shaped so it can't quietly become one:

• robots.txt is not our gate. The RFC 9309 checker ships built in, but a real-time lookup layer isn't a crawler and doesn't do crawler etiquette by default; operators who want it set AGINXBROWSER_HONOR_ROBOTS=1.
• No site-walking API. There is no crawl endpoint and no link-following recursion — every page load happens because an agent asked for that page.
• Built-in budgets. Per-domain: 20 pages/minute. Per interactive session: 200 pages. Toggled via AGINXBROWSER_DOMAIN_RATE_PER_MIN / AGINXBROWSER_SESSION_PAGE_LIMIT (0 disables on your own instance). Generous for an agent grinding through docs or a console; fatal to the page-after-page crawl pattern, including subdomain rotation (one registrable domain, one budget).
• The hosted instance (browser.aginx.net) runs tighter budgets. Every user shares one egress IP, and keeping sites comfortable with that IP is part of the service. Self-host if you want different numbers.
• Need to bulk-crawl a site? Use a crawler. This isn't one, and it won't become one.

What It's For

Not demos — real jobs agent browsers are doing today:

• Grind through admin consoles — AWS / App Store Connect / Google Play, dozens of menu layers per task. Let the agent click; it comes back only when authorization is needed.
• Batch actions behind login — fill carts, dig through order history, check pages that only render while logged in. Inject cookies, operate, export for reuse.
• Past anti-bot walls — Cloudflare protection, Turnstile challenges, TLS fingerprint checks. Stealth mode pushes through instead of retreating at 403.
• The Chinese internet — Baidu / Sogou / WeChat meta-search across 5 engines, correct Chinese page rendering. Not English-web-only.
• On-the-spot scripting — agent reads the page, writes JS, evals it: highlighted comparison tables, reflowed content, product filters on hidden parameters. GreaseMonkey-on-steroids.
• Multimodal vision — screenshots as visual input for look-and-judge flows: picking seats, recognizing layouts, verifying rendering.

Where It Sits in the Computer-Use Stack

Computer-use agents come in two layers. GUI-layer stacks (Cua, desktop CUA agents) drive a whole machine: screenshots of a display in, X11 mouse/keyboard events out, a VM or container per session. Engine-layer browsers skip the desktop — the page itself is the machine. AginxBrowser is the engine layer:

Validated against the GUI layer's own harness: the cua-bench webtop gym — a window manager where every "window" is an iframe — runs on AginxBrowser through the CDP bridge (click task: PASS end-to-end). That's the path that exercises iframe hit-testing, frame-scoped script state, and coordinate translation, which is exactly where engine-layer substitutes usually break. browser-use attaches the same way Playwright does: point cdp_url at the bridge.

Quick Start

Try the hosted instance first: https://browser.aginx.net/

One-command full install (SKILL.md trigger surface + MCP tools + verification):

# Download -> inspect the contents -> run only after review (never blind-run network scripts)
curl -fsSL https://raw.githubusercontent.com/yinnho/aginxbrowser/main/skill.sh -o skill.sh
less skill.sh
bash skill.sh

Register MCP only:

claude mcp add aginxbrowser --transport http https://browser.aginx.net/mcp

Install the skill trigger surface via skills.sh:

npx skills add yinnho/aginxbrowser

Or just hand this block to your AI agent — it sets itself up:

Set up aginxbrowser as your web access layer:
1. Register the MCP server (skip if already present):
   claude mcp add aginxbrowser --transport http https://browser.aginx.net/mcp
   No claude CLI? The HTTP API works directly — no MCP needed.
2. Verify capabilities at https://browser.aginx.net/doctor before calling
   feature-gated endpoints (screenshot/stealth must be listed).
3. From now on, route web tasks through it instead of raw curl:
   - read a page / scrape a URL  -> fetch {url}              (markdown output, Cloudflare bypassed by default)
   - search the web              -> search {q, fetch_top:3}   (5 engines incl. Baidu/WeChat)
   - see a page                  -> screenshot {url}
   - login / form / click-through-> session_create -> session_state -> session_input/session_click -> session_close

Self-hosting:

# macOS / Linux via Homebrew
brew install yinnho/aginxbrowser/aginxbrowser
aginxbrowser doctor   # features + fonts + egress self-check

# Docker (Docker Hub, mirrored on GHCR)
docker run -p 8089:8089 yinnho/aginxbrowser:latest
# (or ghcr.io/yinnho/aginxbrowser:latest)

# Or the prebuilt binary (platform detect + sha256 + mirror fallback + doctor self-check)
# macOS / Linux / Windows (git-bash; prebuilt Windows ships from v0.3.1, full `stealth`+`screenshot` feature set from v0.4.0)
# Cautious: download -> inspect -> run (never blind-run network scripts)
curl -fsSL https://browser.aginx.net/install.sh -o install.sh
less install.sh && bash install.sh
# Or straight in, if you trust the repo:
#   curl -fsSL https://browser.aginx.net/install.sh | sh
# GitHub slow/blocked? AGINXBROWSER_GH_PROXY=https://ghfast.top/ bash install.sh
aginxbrowser doctor   # features + fonts + egress self-check

# Or build from source (--features stealth,screenshot or you lose both)
cargo build --release --features stealth,screenshot

# Start the service
./target/release/aginxbrowser
# → Listening on 0.0.0.0:8089

# Verify
curl http://127.0.0.1:8089/health
# → {"status":"ok","engine":"diting"}

# Fetch a page
curl -sS -X POST http://127.0.0.1:8089/fetch \
  -H "Content-Type: application/json" \
  -d '{"url":"https://example.com"}'

# Search (snippets only by default; fetch_top grabs page bodies for the top N)
curl -sS -X POST http://127.0.0.1:8089/search \
  -H "Content-Type: application/json" \
  -d '{"q":"macbook price","max_results":5,"fetch_top":2,"max_chars_per":2000}'

# Create an interactive session
curl -sS -X POST http://127.0.0.1:8089/session/create \
  -H "Content-Type: application/json" \
  -d '{"url":"https://example.com"}'
# → {"session_id":"s_1","url":"https://example.com/"}

# MCP mode (for AI agents)
./target/release/aginxbrowser --mcp

REST Routes

Every capability is plain HTTP — no SDK required. There is no /openapi.json (the routes are few enough to list here); full request/response fields for each route are in docs/API.md.

Two MCP tools have no REST route: cache (query the local fetch/search cache) and render_markdown (markdown → deterministic HTML artifact). Everything an agent does through MCP is callable over REST above.

Project Layout

aginxbrowser/
├── Cargo.toml
├── build.rs              # V8 snapshot generation
├── js/
│   └── bootstrap.js      # V8 bootstrap script
├── workflow/            # Flow assets: <name>/flow.json replayed by flow_run (drop-in, no rebuild)
├── README.md
├── docs/
│   ├── API.md            # Full API reference (HTTP + MCP)
│   └── integrations.md   # CDP bridge: Playwright / Puppeteer / browser-use
├── bench/                # Benchmark harness + results (vs headless Chrome)
│   ├── README.md         #   methodology + numbers
│   ├── pages.txt         #   fixed 20-page set
│   ├── run.py            #   harness
│   ├── summarize.py      #   TSV → results table
│   └── results/          #   raw run data
└── src/
    ├── main.rs              # HTTP service entry & routing
    ├── server.rs            # Business layer (fetch/click/eval/search)
    ├── session.rs           # Interactive browser sessions
    ├── mcp.rs               # MCP server (37 tools)
    ├── docgen/              # Document layer: markdown → deterministic HTML + inline-SVG diagrams
    ├── render.rs            # Tiered rendering (HTTP direct → diting browser engine)
    ├── store.rs             # Local fetch/search cache (SQLite FTS5, drift hashes)
    ├── download.rs          # Streaming file download (sha256, resume)
    ├── robots.rs            # RFC 9309 robots.txt checker (opt-in gate)
    ├── rate.rs              # Per-domain + per-session budgets
    ├── captcha.rs           # CAPTCHA detection & auto-solve
    ├── firecrawl_compat.rs  # Firecrawl-compatible /v1/scrape endpoint
    ├── video.rs             # Timeline video pump (__timelines seek → ffmpeg → MP4)
    ├── pages.rs             # Page pump (print/slides pagination → PDF/PNG)
    ├── ooxml.rs             # OOXML containers (image-based PPTX/DOCX, stored-ZIP writer)
    ├── diting_cdp/          # CDP bridge (DevTools HTTP + WebSocket)
    ├── doctor_cli.rs        # `aginxbrowser doctor` self-check
    ├── browser.rs           # Top-level API: Browser, BrowserBuilder
    ├── page.rs              # Top-level API: Page, Element
    ├── config.rs            # BrowserConfig
    ├── cookie.rs            # CookieStore
    ├── error.rs             # Error types
    ├── search/              # 15 native search engines, 7 categories
    │   ├── mod.rs           #   SearchEngine trait, Registry, merge/dedupe, progressive backoff
    │   ├── baidu.rs         #   Baidu (JSON API, wreq stealth)
    │   ├── baidu_images.rs  #   Baidu Images (acjson API, images category)
    │   ├── bing.rs          #   Bing (HTML parsing, plain reqwest)
    │   ├── bing_images.rs   #   Bing Images (images/async endpoint, images category)
    │   ├── bing_news.rs     #   Bing News RSS (news category; proxy-first)
    │   ├── sogou.rs         #   Sogou web (HTML parsing, plain reqwest)
    │   ├── sogou_wechat.rs  #   Sogou WeChat (HTML parsing + /link resolution)
    │   ├── duckduckgo.rs    #   DuckDuckGo (html.duckduckgo.com, general; direct-first)
    │   ├── google.rs        #   Google (HTML parsing, wreq stealth + proxy)
    │   ├── stackexchange.rs #   Stack Overflow (SE API v2.3, code category)
    │   ├── github_repos.rs  #   GitHub repos (api.github.com, code category)
    │   ├── arxiv.rs         #   arXiv (Atom API, academic category)
    │   ├── huggingface.rs   #   HF Hub models/datasets/spaces (ai category)
    │   ├── npm.rs           #   npm packages (npms.io API, packages category)
    │   ├── pypi.rs          #   PyPI name resolution (JSON API, packages)
    │   └── meilisearch.rs   #   Private-index adapter (env-configured)
    │
    ├── diting_dom/          # HTML parsing, DOM tree, CSS selectors
    ├── diting_css/          # CSS parsing + cascade
    ├── diting_net/          # HTTP client, cookies, encoding, proxies
    ├── diting_js/           # V8 runtime, JS ops, module loading
    ├── diting_layout/       # Taffy-based layout, floats, hit-testing
    ├── diting_fonts/        # Bundled CJK font subset, fallback
    └── diting_browser/      # Page navigation, lifecycle, browser context

Build

# Standard build (no stealth; TLS fingerprint features inactive)
cargo build --release

# With stealth (requires go + cmake + C++ toolchain; enables TLS fingerprint spoofing)
cargo build --release --features stealth

# With screenshot rendering (enables /screenshot; adds the rendering stack, +30-40MB)
cargo build --release --features screenshot

# Full featured (recommended for production)
cargo build --release --features stealth,screenshot

Requirements: Rust 1.78+; the V8 static library downloads automatically on first build. The stealth feature additionally needs go, cmake, and a C++ compiler. The screenshot feature ships with a bundled CJK font subset (GB2312 + common symbols) — no system fonts required for correct Chinese rendering.

If your network can't reach the rusty_v8 CDN (build hangs with zero progress after "downloading v8"), pre-fill ~/.cache/rusty_v8 with the librusty_v8.a.gz for your version (fetch it from any reachable mirror/host and gunzip into place) and the build script skips the download.

Runtime Environment Variables

API Documentation

Full API reference → docs/API.md CDP integration guide → docs/integrations.md — Playwright / Puppeteer / browser-use one-liners Security audit notes → docs/skills-sh-audit.md — why skills.sh shows "Critical Risk", and which real product feature each warning corresponds to

Covers:

• All 37 HTTP endpoints (/fetch, /search, /screenshot, /video, /pdf, /download, /v1/scrape, /flow/run, /doctor, 19 session endpoints, CDP discovery, MCP transport)
• All 37 MCP server tools and their parameters
• Claude Code / Claude Desktop / Cursor client configuration
• Environment variables, error codes, per-site scraping examples

Plugging Into Other Systems

AginxBrowser is pure attach-alongside infrastructure — like a real browser, it runs as an independent service that anything can call, without embedding host code or polluting host config. Deploy one instance per machine (under systemd) and every app needing "render + scrape" capability shares it.

Three attach points:

• HTTP — /fetch, /search, /screenshot, /download for any language with an HTTP client
• MCP — one line into Claude Code / Cursor / Claude Desktop (above)
• CDP — point Playwright / Puppeteer / browser-use at ws://your-host:8089/devtools/browser/<id>; agent-browser drives it with --cdp (--cdp-port N binds loopback for that). Google's chrome-devtools-mcp attaches too: aginxbrowser --cdp-port 9223 + chrome-devtools-mcp --browser-url http://127.0.0.1:9223 — its full tool battery (pages, snapshot, evaluate, screenshot, network, resize) passes against the diting engine. See docs/integrations.md

Integration: read the environment variable AGINXBROWSER_URL=http://127.0.0.1:8089. Unset → behavior unchanged; set → risk-controlled sites automatically route through AginxBrowser for rendering, falling back gracefully on failure.

Known Limitations

1. Screenshots are opt-in: /screenshot requires cargo build --release --features screenshot (adds the diting rendering stack). The default (and only) render engine in that build is diting — our own CSS+layout+paint stack, zero Blitz/Stylo code. The pinned-rev Blitz reference pipeline is a separate opt-in, --features blitz-reference, for comparison renders and the dual-engine cross-check tests. Complex-site CSS is approximate on both (not pixel-perfect like Chromium)
2. Element coordinates supported: /screenshot with selector returns element page coordinates (selector_rects, CSS px); selector alone crops directly to that element. Inline elements (<a>text</a>) get a rect too on the default diting engine — a union of their flattened inline content, strut-expanded to the element's own line-height like Chrome reports for replaced-only inlines (<a><img></a> → line-box height, not the image height). Empty inlines still have no rect — pick a block ancestor there
3. JS interaction broadly works; heavy-fingerprint pages may still fail: React/Vue event delegation works normally (URL-reflection attributes like src/href resolve to absolute URLs so Next.js/webpack hydrate and clicks trigger handlers). Heavy-fingerprint auth pages (WorkOS/Cloudflare) probing navigator.plugins, WebGL canvas etc. may still break until stealth fingerprint coverage completes
4. Proxy support: HTTP/HTTPS/SOCKS5 via AGINXBROWSER_PROXY
5. Hard risk-controlled sites: Baidu Wenku unsupported; Zhihu articles need a valid __zse_ck

Star History

If AginxBrowser saved you a headless-Chrome fleet or a scraping headache, a star is how other agents (and their humans) find the project.

License

Apache-2.0.