Repository navigation
n8n Integration
n8n-nodes-navette — navette in your n8n workflows, as a community node. Navigate, read, screenshot, click, type and evaluate a real system WebView (WebKit on macOS, WebView2 on Windows, WebKitGTK on Linux) driven by the ~1 MB navette daemon — no Chromium download, no 557 MB browser cache in your CI or container.
Source: integrations/n8n-nodes-navette/ · npm package: n8n-nodes-navette · in-repo version: 0.3.1 · zero dependencies by design.
Beginner — build the flow by hand. Every action speaks plain names: Open a page → Take a screenshot, Open a page → Read the page → send to Slack. The Navigate action carries an Options section (window size, wait-for-element, screenshot) so a full capture is ONE node.
The "go to this site and extract this" way — AI Agent. The node is usableAsTool: drop an AI Agent node (any chat model), add the navette actions as its tools, and just tell it what you want:
"Go to https://news.ycombinator.com and give me the top 5 titles with their points."
The agent opens pages, reads the markdown it gets back, clicks if it must, and answers — no workflow wiring beyond the two nodes. Pair with n8n's built-in Information Extractor for strict-JSON answers. Version 0.3.1 adds the dedicated agent operation ("Browse for me (AI)") — the self-driving operation with Navigate Options all-in-one.
| Operation | Body | Note |
|---|---|---|
| Navigate | url, withContent, format | redirects complete fine |
| Read | format | markdown for LLM steps |
| Screenshot | — | PNG as binary output (attach to email/Slack directly) |
| Click | selector, waitNavigation | pointer+mouse events — Radix menus open |
| Type | selector, value | React-safe native setter |
| Evaluate | js | escape hatch |
| Wait | selector, ms | polls until it exists |
| Viewport | width × height | per session |
| Login | site | keychain-bound, origin-bound fill — the workflow never sees the password |
| Export / Import State | — | the cookie jar out and back in (storageState equivalent) |
| Sessions / Close | — | housekeeping |
| Agent (0.3.1+) | mission / startUrl | the "Browse for me (AI)" self-driving operation |
Sessions are named contexts inside the daemon — one per site; cookies persist between nodes and runs.
# the daemon first (brew / cargo / docker — one binary)
navette serve --port 8765In n8n: Settings → Community nodes → Install → n8n-nodes-navette. Then create the navette daemon credential (Base URL http://127.0.0.1:8765, Token only if the daemon runs --token).
Local install for development:
cd integrations/n8n-nodes-navette && npm pack # zero dependencies
docker cp navette-nodes-navette-<ver>.tgz n8n:/tmp/
docker exec -it n8n npm install /tmp/navette-nodes-navette-<ver>.tgz
docker restart n8nDocker networking: if n8n runs in Docker, the host daemon is reachable at http://host.docker.internal:8765 — which requires the daemon to accept that Host header:
navette serve --port 8765 --allow-host host.docker.internal # pair with --token beyond a private LAN-
navette creds set mysiteonce (password → OS keychain, bound to the origin) — then the Login node logs in without the workflow ever seeing the password; or log in by hand once and Export State → Import State the cookie jar between runs. - Navigate → Type → Click the real UI; Read/Screenshot to verify.
- Same daemon, same sessions, ~1 ms actions — the measured Playwright comparison is in
demo/wordpress(see Demos).
-
examples/wordpress-audit-workflow.json— the WordPress audit: 9 of 12 operations in one realistic run. -
examples/devto-replybot-workflow.json— the dev.to reply bot.
The node wraps a loopback JSON API — there is no SDK to vendor and no browser to download, which is the whole point. package.json ships an empty dependencies object on purpose; keep it that way. (Compatibility note: package.json declares both n8n.nodes and the legacy n8n-nodes key — n8n-core 2.x reads n8n.nodes.)
slabbdev/navette · v1.9.0 · MIT — the browser for agents. One tiny Rust binary, the WebView your OS already ships.
navette — the browser for agents
- Home
- Use it
- Sessions, secrets, humans
- Ecosystem
- Evidence
- Project