Changelog

​

Features

​

API trigger endpoint now accepts bodyless requests

​

Full endpoint URL and method in the API trigger panel

​

Features

​

Opt-in CloakBrowser browser engine

• Default (USE_CLOAK_ENGINE unset or false): Playwright Chromium patched at the JS level with playwright-extra and puppeteer-extra-plugin-stealth. Same behavior as previous releases.
• Opt-in (USE_CLOAK_ENGINE=true): the CloakBrowser stealth-patched Chromium binary. Evasions are applied at the binary level rather than injected at runtime, which is harder for modern bot detectors to fingerprint. Use this when the default stack is being detected on a target site — typically CAPTCHA walls that trigger on Playwright/Chromium runtime tells.
• License key: set CLOAKBROWSER_LICENSE_KEY to unlock the latest stealth binary. Without a key, CloakBrowser falls back to the free legacy binary. You can also run npx cloakbrowser login to write the key to ~/.cloakbrowser/license.key, which cloakbrowser reads natively.
• Lean installs: the CloakBrowser binary is only fetched during postinstall when the flag is enabled, so the default install stays the same size as before.

USE_CLOAK_ENGINE=true
CLOAKBROWSER_LICENSE_KEY=your_license_key_here

​

Automatic cookie-consent dismissal in Agent mode

• Covers hundreds of providers (OneTrust, Cookiebot, TrustArc, Quantcast, and more). The script always clicks reject or close — never accept.
• Always on in Agent mode. Headful mode is unchanged (it’s for human interaction), and Scrape mode doesn’t render banners because it no longer launches a browser.
• If no matching banner is present, the script silently no-ops so your task runs as usual. If a specific site’s dialog isn’t covered, you can still author an explicit Click block.

​

Changes

​

Scrape mode is now browserless (no screenshots or video)

• Faster and cheaper: No browser process means dramatically lower CPU, memory, and cold-start cost per scrape.
• Preserved: Selector extraction, HTML cleanup, link extraction, proxy and user-agent rotation, the headful → scrape cookie handoff, and the sandboxed extraction-worker script pipeline all still work exactly as before.
• Removed: No screenshots and no video recordings in Scrape mode. There is no page to photograph. The Results panel’s Screenshot pane now shows “Scrape mode does not support screenshots” instead of the misleading “Waiting for Frame…” placeholder. If you need visual evidence of the page, run the task in Agent mode.

​

Features

​

Visual field-mapping mode for extraction scripts

• The mode toggle at the top of the editor flips between Visual and JavaScript. The selected mode now has a solid fill so it’s obvious which one is active.
• Every field has a target icon that hands off to the Headful Browser inspector. Click an element on the page and the selector, plus a row of alternative candidates, get written back into the field.
• Playwright-only :has-text(...) selector candidates are filtered out of the pick list for extraction fields only. Extraction scripts run through native document.querySelector, which does not understand that pseudo-selector, so surfacing it would have produced selectors that returned null at extraction time. Action blocks (Click, Type, etc.) still receive the full candidate list because Playwright runs those.
• Switching to JavaScript mode surfaces the auto-generated script in the code editor, still fully editable. Existing hand-written extraction scripts open in JavaScript mode automatically so your code is never hidden behind an empty field list.
• The canvas Extraction Script modal no longer closes when you click the backdrop, so a stray click on the dimmed area outside the modal will not throw away in-progress edits. Use the Done button or the close icon in the header to save and dismiss.

​

Updates

​

Agent Mode description clarified in Task Settings

• The label under Agent Mode now reads “Custom action sequence with logic.”
• No behavior change. If you were choosing Scraper Mode because you didn’t want the runner making its own decisions, you can safely use Agent Mode — the same author-defined blocks execute in the same order on every run.
• If you want an actual AI-driven, decision-making runner, that lives in MCP Integration, not in Agent Mode.

​

Updates

​

headless flag and automatic fallback for POST /api/browser/open

• Pass "headless": true in the request body to launch Chromium without a display. This is the right choice when you run Figranium directly on a Mac (outside Docker) or on a headless CI runner. The HEADLESS environment variable is also honored.
• If you don’t set headless and the initial headful launch fails because there is no display, Figranium retries once in headless mode and returns the running session. External orchestrators no longer have to detect HEADFUL_DISPLAY_UNAVAILABLE and reissue the call themselves.
• The display-unavailable detector now also matches no display server, X11 connection failed, cannot open display, and unexpected target closed errors during launch, so the fallback triggers on Mac hosts where the underlying error text differs from Linux.

curl -X POST http://localhost:11345/api/browser/open \
  -H "x-api-key: $FIGRANIUM_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"url": "https://example.com", "headless": true}'

​

Bug fixes

• Headful VNC viewer no longer gets stuck on “Reconnecting…” — start-vnc.sh was embedding literal double-quote characters into the x11vnc -passwd argument because of unquoted shell word-splitting, so the password x11vnc actually enforced never matched the one served to the noVNC client. Every connection attempt failed authentication and dropped immediately, which showed up in the UI as an endless reconnect loop. The password is now passed through a bash array as a single unmodified argument. If you were affected, pull the latest image and reopen the headful session — no config change is required.
• Proxy credentials pasted as a full URL are now split out on add — Pasting http://user:pass@host:port into Settings > Proxies > Add Proxy > Server previously stored the whole URL (credentials included) as the server value, which surfaced as a duplicate-looking entry in the list and failed upstream authentication in rotation pools. normalizeProxy now extracts embedded credentials into username and password on both the add and bulk-import paths. Existing broken entries self-heal on next read — no manual cleanup needed. See Proxy Rotation — Adding a proxy.
• Backdrop blur restored on modals and overlay panels in Dark and Solarized Dark themes — A dark-theme CSS override was force-replacing translucent bg-black/NN backgrounds with a fully opaque color, defeating backdrop-blur on the Task Settings panel, the confirm modal, and other overlays. The override has been removed, and the confirm modal now uses a translucent theme-aware background instead of a hardcoded opaque one.
• Browser open and inspector highlight on Mac — Launching the managed browser session from POST /api/browser/open and resolving a targetHint through POST /api/inspector/highlight now work when Figranium runs on macOS. Chromium is launched without the --disable-gpu, --window-position=0,0, and --start-maximized flags on Darwin, which were causing the process to exit before the first page was ready. The CDP Browser.setWindowBounds maximize call is also skipped on Darwin. Selector generation, XPath resolution, and the top-match highlight overlay for /api/inspector/highlight now run in a single in-page evaluation, so they don’t lose the element handle to a mid-flight page navigation.
• GitHub star prompt no longer appears on first run — The in-app prompt asking you to star the Figranium repo is now gated on both 3+ successful task runs and 3 days since your first run (tracked locally in the browser). It also uses a quick slide-in animation instead of the previous fade so it’s easier to notice when it does appear. Dismissing or clicking through still permanently silences it.

​

Updates

​

Single self-theming brand icon

• The mark fills black by default and switches to white inside @media (prefers-color-scheme: dark), so it stays legible on both light and dark surfaces.
• Use this file when you can only supply one logo URL — for example, third-party integrations, README embeds, or partner directories that accept a single <img src>.
• No configuration change is required. Existing embeds keep working; you can drop the separate dark/light variants from your integration if you were maintaining both.

<img src="https://<your-figranium-host>/figranium_icon.svg" alt="Figranium" width="32" height="32" />

​

Features

​

Multi-theme support

• Switch themes from Settings → System → Theme. Each option shows a preview image plus a surface and accent color swatch.
• The choice is stored in the browser’s local storage (figranium.theme), so it persists across reloads on the same device.
• The first time you open Figranium after upgrading, a one-time picker prompts you to choose a theme. Dismissing it also persists locally and it will not appear again.
• Themes are implemented as CSS custom properties (--app-bg, --app-surface, --app-accent, and the full syntax-highlight palette), so the whole UI — including code blocks — retunes for readability under each background.

​

Authenticated programmatic browser & inspector API

• POST /api/browser/open launches or reattaches a managed headful session and returns { sessionId, status, wsEndpoint }.
• POST /api/inspector/highlight activates the inspect overlay on the active session and, given a targetHint, returns up to five candidate elements with CSS selectors, XPath, confidence scores, and an optional base64 JPEG snapshot of the viewport.
• PATCH /api/tasks/:id performs a partial update on a task. Figranium snapshots the current task into the version history before applying the change.
• DELETE /api/tasks/:id now also removes any in-process schedule registered for the task, so a scheduled run cannot fire after deletion.

curl -X POST http://localhost:11345/api/inspector/highlight \
  -H "x-api-key: $FIGRANIUM_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"url": "https://example.com", "targetHint": "Sign in"}'

​

Security

​

VNC and websockify path hardening

• The noVNC/websockify proxy path now requires authentication before it will accept an upgrade or serve public/novnc.html. Unauthenticated attempts are dropped.
• websockify binds to the IPv4 loopback (127.0.0.1) explicitly, closing a mismatch where Docker on macOS could bind unpredictably across IPv4/IPv6 loopback and leave the VNC stream unreachable or unexpectedly reachable.
• sessionId and taskId are strictly validated before they are used to look up files on disk, closing a path-traversal edge case.
• Raw cron ranges are now validated for correct ordering as well as bounds, so a malformed field cannot slip past the scheduler.

​

Bug fixes

• Captures now surface reliably — Runs no longer produce screenshots or recordings that fail to appear in the UI. The capture read path in server.js and src/server/routes/data.js now matches the actual write path, and captures are also served from the alternate public/captures location used inside Docker.
• Captures survive container restarts — Runtime capture artifacts persist to a host volume instead of being lost when the container is recreated.
• Persistent browser session ID — Session IDs are stored across reconnects instead of being regenerated on every new session, so external tooling that pins to a sessionId stays valid.
• Mac/Docker WebSocket disconnects — Loopback WebSocket connections used by the headful viewer no longer drop on Mac hosts, and the HeadfulModal no longer flashes its loading state during a normal reconnect.
• Apple Silicon headful browser “Connecting… Disconnected” loop — The GHCR publish workflow now builds and pushes a true multi-architecture image (linux/amd64 and linux/arm64). Apple Silicon Macs pull the native ARM64 image instead of running the AMD64 image under Docker Desktop’s QEMU emulation, where Xvfb and x11vnc were prone to crashing seconds after start. That crash also silently broke the selector-picker SSE stream, which shares the same headful browser session.
• Headful viewer auto-reconnect — Xvfb, x11vnc, and websockify each now run inside restart loops (with per-process logs under data/xvfb.log, data/x11vnc.log, and data/novnc.log), and the noVNC page auto-reconnects with backoff on RFB disconnect. A quick post-connect drop is now surfaced as “Browser session crashed, retrying…” instead of a plain “Reconnecting…” so a genuine crash is visually distinguishable from a normal reconnect.
• Theme contrast fixes — Dot-grid canvas background, theme-accent buttons, and previously hardcoded blue accent text (API keys, Add API Key button) are readable across all four themes; syntax-highlight colors were retuned to meet WCAG AA contrast on every theme’s actual code background.

​

Updates

​

Exported task filename

​

Font change

​

Under the hood: CAPTCHA detection scaffolding

​

Features

​

Model Context Protocol (MCP) server

• Prebuilt image: pull ghcr.io/figranium/figranium-mcp:latest — no Node.js install or local clone required.
• STDIO transport: works with any MCP-compatible client over the standard STDIO transport.
• Point at your instance: configure FIGRANIUM_BASE_URL and FIGRANIUM_API_KEY in the client’s MCP server config. Use http://host.docker.internal:11345 when Figranium runs on the same host.
• Registry namespace: clients that support automatic registry resolution can install it as io.github.figranium/figranium-mcp.

​

Features

​

Named session snapshots (sessionId)

• Pass sessionId: "acct-alice" on a run and Figranium loads data/sessions/acct-alice.json as the browser’s initial storage state. After the run, the current state is written back to the same file.
• Different sessionId values give you fully isolated logged-in identities on the same target — one per customer, tenant, or account, without touching the shared profile directories.
• sessionId is sanitized to [a-zA-Z0-9_-]; other characters are stripped for path safety.
• Combining sessionId with statelessExecution: true keeps the run stateless — the snapshot is not written.

​

Updates

​

Async and top-level return in Run JavaScript

​

Security

​

Hardened VNC and websockify access

• Localhost-only binding: both x11vnc (port 5900) and websockify/noVNC (NOVNC_PORT, default 54311) now listen on 127.0.0.1 only. Publishing these ports on the Docker host no longer exposes them.
• Password-protected VNC: x11vnc requires a random password generated on first start and stored at data/vnc_password.txt. The embedded viewer fetches it automatically from the new authenticated endpoint GET /api/headful/vnc-password.
• Authenticated /websockify upgrades: every WebSocket upgrade to /websockify on the main Figranium port must pass the IP allowlist, an Origin/Host match (CSWSH protection), and present either a session cookie or an API key. Unauthenticated attempts are dropped and logged.

​

Updates

​

Expanded PostgreSQL storage

• More data in Postgres: proxy configuration, saved credentials, AI model selections, and Ollama API keys are now persisted in the database when DB_TYPE=postgres.
• SSL support: a new DB_POSTGRESDB_SSL=true environment variable enables encrypted connections to managed Postgres providers (RDS, Cloud SQL, Supabase, Neon, etc.).
• Longer API keys: API key columns are now TEXT instead of VARCHAR(255), and existing tables are migrated automatically on startup.
• Graceful fallback: if the database is unreachable at startup, Figranium falls back to file-based storage.

​

Expanded SSRF protection

• More IPv4 ranges blocked by default: in addition to RFC 1918 private ranges and loopback, Figranium now blocks IETF protocol assignments (192.0.0.0/24), TEST-NET ranges, benchmarking (198.18.0.0/15), shared CGN space (100.64.0.0/10), multicast, and other reserved space.
• Full IPv6 coverage: loopback (::1/128), unique local (fc00::/7), link-local (fe80::/10), unspecified, and multicast ranges are blocked.
• Hostname blocking: localhost, *.localhost, and host.docker.internal are blocked unless ALLOW_PRIVATE_NETWORKS=true.
• Proxy server validation: proxy URLs added through Settings or the API are validated against the same blocklist. Invalid entries are rejected with INVALID_URL, and bulk imports fail atomically if any entry is unsafe.
• Ollama URL validation: Ollama base URLs are validated both at save time and again at request time, with every redirect hop re-checked and sensitive headers stripped on cross-origin redirects.
• Output provider credentials: baseUrl values (e.g. Baserow) are validated when credentials are saved, rejecting unsafe URLs with INVALID_BASE_URL.
• Redirect protection: outbound webhook and output provider requests now validate every hop in an HTTP 3xx chain (up to 5 redirects).

​

Gemini API key transport