Changelog

Changed

• Improved the titles and dialog captions generated during navigator sessions.

Added

• Navigator video captions can now be placed at a configurable position on the frame.
• Caption display time now adjusts to how long the caption takes to read.

Changed

• Navigator video captions are now drawn on top of the captured frames, with improved font rendering and sizing.
• The bundled agent navigation guide now covers navigator captions.

Fixed

• Wait steps in navigator videos are now always recorded, even when the page doesn’t visibly change during the wait.

Added

• Navigator now supports System One’s image protocol when using jev.

Changed

• Navigator video captures now skip runs of identical frames, so recordings no longer stall on unchanged screens.

Added

• Recorded navigator videos (WebM/MP4) can now show the mouse cursor, so you can see pointer movement and clicks.
• zshot now gives AI agents that run it basic guidance on how and when to use it.

Changed

• Updated the bundled agent skill and its navigation reference.

Added

• New flag-docs-json option that outputs help for every flag as JSON.
• Output file types now accept the same aliases on the command line as in JSON requests.

Changed

• Storage import now detects which browser a profile belongs to and imports data for the right sites.
• Text output formats now write to stdout when no output filename is given.
• Navigator-only options now cause an error when used without navigation, instead of being silently accepted.
• A warning is now shown when Chromium language/international preferences such as accept_languages are set but will not take effect.
• OpenRouter navigator connections now default to the JEV model when a decision protocol is selected.
• Video selector capture is now subject to license checks, like other video features.
• Simplified the Kubernetes and Docker Compose deployment examples.
• Release binaries are now stripped, reducing download size.

Fixed

• The navigator now handles focus changes better while typing, returns to the intended field, and verifies that input was actually entered.
• Generated element selectors now prefer meaningful class names over purely structural ones.
• The navigator can now find and scroll content inside inner scrollable panels, and reports their contents correctly.
• The navigator now counts text changes and scrolling as page progress, so it no longer reports a stall too early.
• The navigator is now told when one of its actions is refused, so it can choose another approach.
• Credential values entered by the navigator no longer appear in page content sent to the model afterward.
• Exported Puppeteer and Playwright replay scripts now include the final navigation and check that the run ends on the expected page.
• In CDP mode, page targets now appear in the expected order.

Changed

• Updated skills documentation

Added

• Openrouter provider
• Added navigator-data for providing job specific parameters to a task

Changed

• Added ability for navigators to proceed past errors with reload and more error context information

Added

• CDP mode for replaying CDP scripts
• Playwright and puppeteer recording of navigator actions
• Cloudflare navigator compatibility
• Enhanced actions support for typing, keydown and keyup
• Automatic parallelization of crawlers

Changed

• Many internal subsystems reworked
• How progress is recorded for navigators from screen to screen

Fixed

• Frames reporting progress when no paint had been received

Added

• NOTE command for navigator memory
• Support for storage, from existing browsers or JSON file
• Added scroll commands for navigators

Changed

• Navigator videos do not record intervals when thinking

Fixed

• Visibility issue for AX tree visibility

Changed

• Improved memory in several navigator usage cases
• Updated CEF
• Added token outputs to prometheus metrics

Changed

• Fix diagnostics reproductions
• Adjust command handling for navigators

Changed

• Improved navigator composition through supporting multiple navigator-types over a “generative” type
• Allow specifying a navigator to take on generative tasks through a write=true parameter
• Allow specifying when secondary navigators should be used via the when= argument
• Allow navigators to failover toe ach other if one becomes unavailable
• Support mercury ( inception labs )
• Improve prompt efficiency, reducing number of tokens and calls
• Improve mouse usage
• Add –navigator-usage, to output total token usage after a run
• Improve prompting on several navigator actions, to better hint on routing

Changed

• All emulation options now free
• Updated CEF
• Added –trial command to reach trial signup

Changed

• Improved navigator decision making capabilities

Changed

• Update CEF.
• Disabled background model acquisition to prevent unwanted network activity during captures.

Removed

• Removed several unused command-line flags.

Added

• New web bot auth support, letting zshot present verifiable signed credentials to sites that require them.
• Navigator now supports pointer down/up events and dragging items, in addition to clicks.

Changed

• Refreshed the default LLM/chat provider models.

Added

• Navigators can now emit a goal output, letting you capture the result of a navigation run.
• New option to limit how many times a navigation action can repeat, preventing agents from getting stuck in loops.
• Support for bearer token authentication with Anthropic, automatically selected based on your token.

Changed

• Smooth scrolling is now disabled by default during navigation.
• CDP operations no longer block other render requests, so concurrent captures run more reliably.
• Updated CEF.

Fixed

• Prevented crashes caused by writes to stdout/stderr that could abruptly terminate the process.

Added

• Support a generative navigator type, allowing you to combine routing-only and generating backends

Added

• Set arbitrary Chromium preferences from the command line.

Changed

• Refreshed browser flags and preferences to match the current Chrome/CEF version.
• Improved automated element activation to skip text fields when using the non-generative decider.
• More accurate browser emulation

Added

• Navigator can now handle more pages, moving faster and skipping forms it can’t work with.

Changed

• Added more context for navigator history

Added

• New Typesafe.ai navigator
• License refresh: refresh an existing license when needed, in addition to registering and releasing.

Changed

• BMP pixel format is now available in the free tier.

Fixed

• Corrected automatic MIME type detection for application/octet-stream inputs.

Added

• Credentials can now be pulled from environment variables instead of being passed directly on the command line.
• You can now set the initial URL scope for the navigator to enforce your security requirements, with a configurable allowlist.

Changed

• Updated the default models for LLM providers.

Added

• The navigator can now handle file upload dialogs, filling them from context instead of requiring direct file-input interaction.
• Multiline text values are now captured and set correctly in text areas during form filling.
• When a link opens a new tab or window, the navigator can switch to it to complete a task.

Changed

• Status output on stderr now respects terminal column widths, with cleaner final output written to stdout.

Added

• URL allowlist and permission rules now support more flexible glob patterns and can be applied with finer control over how they match.

Changed

• Navigation failures now cause the render to fail instead of producing a partial or empty result, so errors surface clearly.

Added

• You can now reference remote files for form uploads — zshot will fetch and store the remote file on demand when a page needs it.
• Navigator credentials can now be supplied from an external process or an HTTP server, in addition to existing methods.

Changed

• Update CEF.

Changed

• Performance traces now run one at a time, preventing conflicts when multiple trace attempts overlap.

Added

• Server now queues requests by default instead of immediately returning HTTP 429 when busy. You can opt back into the old 429 behavior when needed.

Fixed

• Prevented full-page captures from deadlocking when the request queue was full, improving reliability under heavy load.

Changed

• Added memory guidance to the sample Docker Compose and Kubernetes deployment files.

Changed

• Docker image now has font cache correctly baked into layer

Added

• Sandboxed container runs are now supported, including a seccomp profile and a sandbox self-check to verify your environment.
• Published example Docker Compose and Kubernetes deployment files to help you get up and running.

Changed

• Updated the default models for LLM providers.
• Update CEF and Chromium to 152

Fixed

• Shell completion files are now placed correctly in Linux packages, resolving packaging warnings.

Changed

• Refreshed the default models used by the built-in LLM providers.
• Updated the End User License Agreement.

Added

• GPT6 Astra and Gemini 3.8 Flash are the new default models for those providiers
• license-key now updates your config if you specify it without any other arguments

Changed

• JPEG XL output is now available on all license tiers.
• Full-page capture is now available on all license tiers.
• Updated the bundled skill documentation with guidance on archival, uploads, downloads, and credentials.
• Updated CEF.

Fixed

• Fixed an iframe load race that could affect captures on Linux.
• Improved signal handling and cleanup on exit for more reliable shutdown.

Changed

• Refreshed the default models used by the built-in LLM providers.
• Reworked LLM provider backends (OpenAI, Anthropic, Ollama, Gemini) for more consistent behavior across providers.

Removed

• Removed the built-in Swagger UI and its associated server endpoint and options.

Added

• JavaScript dialogs (alert, confirm, prompt, beforeunload) are now automatically dismissed so captures no longer hang on unexpected dialogs.

Fixed

• Device overrides now apply consistently and take effect before capture, producing more reliable results.

Changed

• Update CEF

Changed

• Updated to FFmpeg 9.0

Added

• Container mode has friendlier license activation for server environments

Changed

• Improved license validation reliability, including handling of slow or unresponsive licensing servers.

Added

• Support for the stateless MCP protocol.

Changed

• Update CEF.

Changed

• Update CEF.

Changed

• Update CEF.

Changed

• Update CEF.

Fixed

• Form filling now correctly handles fields that already contain values.

Added

• Downloads triggered via blob: URLs are now captured and made available.

Added

• Navigator can now upload files to web forms during automated sessions.
• New download handling: files triggered by a page (or by navigator actions) can now be captured and saved, including downloads into WARC archives.

Added

• Support text file output
• Provide MIME type hints for input files when the type can’t be inferred automatically.

Changed

• Smoother handling of file inputs on the command line, including improved extension awareness.

Fixed

• Safer handling of MIME types to guard against potentially dangerous content, with data URIs now triggering the correct capture behavior.

Fixed

• Full-page captures now wait for page navigation to complete before capturing, preventing incomplete or blank screenshots.

Added

• More detailed logging about navigation decisions when the navigator or crawler traverses a page graph.

Changed

• License validation failures now fall back to the free tier instead of blocking usage — the app keeps working even if licensing can’t be verified.

Added

• Cache entries can now be invalidated on demand via HTTP DELETE requests to the server.

Changed

• When capturing multiple pages (e.g. from a sitemap), the most specific page limit now takes priority when applying capture budgets.

Changed

• Improved the readability and interpretability of capture status output.

Added

• WACZ output format for capturing web archives in a portable, shareable package.
• CDX index generation for WARC/WACZ captures, making archives easier to index and replay.
• Markdown output can now reference contextually appropriate image URLs or embed image content directly.

Changed

• Enterprise captures now default to using all available CPU cores for parallelism instead of a single worker.
• Server mode is now limited to a concurrency of 1 in Pro mode.
• Output format options are now validated for compatibility, giving clearer errors when incompatible options are combined.
• Licensing restrictions are now enforced more consistently across capture types.
• DNS resolution now always uses the system resolver.

Fixed

• File URL access is now correctly limited to within the specified subdirectory, and temporary files are scoped to their own subdirectories rather than appearing alongside sibling files.

Added

• Server mode can now cache requests and act as a browser CDN, serving repeat captures faster.
• New Prometheus metrics with per-output-type render histograms for easier operational monitoring.

Changed

• Sitemap crawls now expand automatically with a Pro license; without Pro, only the given URL is captured.
• Server endpoints are now controlled by explicit enable flags instead of disable flags.
• WARC captures and crawls now wait for in-flight assets to finish loading before completing, so archives better match what you’d expect to see.
• Device memory now reports values that match the target device (including undefined where appropriate).
• Logging is initialized earlier so warnings during startup are no longer lost.
• Server-only and other context-specific flags now error clearly when used in a mode where they have no effect.
• Markdown output now ensures image URLs are absolute, and images can now have their ocntent embedded into the result Markdown
• Update CEF.

Fixed

• WARC archives now correctly mark truncated response bodies as truncated.
• Diagnostics output now reflects your actual input.
• Fixed a potential crash from invalid string slicing in argument handling.
• Removed a stray error message printed during certain string conversions.

Added

• New Galaxy and iOS device presets for device emulation.

Added

• Crawl mode can now download linked files (e.g. documents) and capture media resources encountered during a crawl, saving them into your WARC archive.
• New option to cap the maximum number of links fetched during a crawl.
• You can now stop a crawl mid-capture and keep the output already collected.

Changed

• Crawl-related flags are now grouped under a consistent crawl namespace instead of being tied to specific output formats. You may need to update existing command lines that used the old flag names.
• WARC and HAR archives are now written as a stream, greatly improving scalability and memory use when capturing large sites.
• Connection strings now infer model configuration automatically, reducing the amount you need to specify by hand.

Added

• The AI navigator can now interact with form controls — select boxes/dropdowns, radio buttons, and checkboxes — when filling out and navigating pages.

Fixed

• Improved handling of cross-domain requests during crawls so custom headers stay within the intended domain boundaries.

Added

• LLM output formatter can now be specified per request, in addition to globally.
• Exposed more capture and navigation options to the JSON/HTTP API.

Changed

• PNG, JPEG, TIFF, and BMP are now available as free formats.
• Additional emulation options are now available in the standard tier.
• Updated CEF.

Fixed

• Navigator video now reliably follows the intended instructions.

Added

• New navigator actions to fill form fields and submit forms.

Changed

• Screenshot capture now uses CDP for more reliable results.
• Updated CEF.

Added

• Docker server image now includes a healthcheck and additional setup conveniences for the entrypoint.

Removed

• Removed the local LLM provider, including bundled model download/management and on-device inference. LLM features now require a remote provider.

Added

• New Docker server container image for running zshot as a long-running service, including a built-in health check.

Added

• Capture entire sites to WARC by following a sitemap, with automatic link discovery for multi-page crawls.
• WARC output is now more compatible with tooling, including revisit records to deduplicate repeated resources across pages.
• New accessibility-tree (axtree) output alongside your captures.
• More lenient handling of JSON request bodies when posting to the HTTP server.

Changed

• Update CEF.
• Improved WARC page records for better parity with other crawling tools.

Fixed

• Resolved a content-encoding mismatch that could break WARC replay verification.

Added

• zshot can now be installed and updated via an official APT repository (Debian/Ubuntu .deb packages).

Added

• New command-line options for diagnosing and tuning captures.
• URL preflighting that checks a page before capture, reducing wasted work on unreachable URLs.

Changed

• Improved PNG render performance via tuned encoding settings.
• Skip favicon fetches when they aren’t needed, speeding up captures.

Fixed

• More reliable scheduling of capture work, improving responsiveness and stability.
• Avoid a spurious failure when the navigation step limit is reached.

Changed

• Adjusted process exit codes

Added

• New --formats argument to list all supported output formats from the CLI.
• Runtime detection of available codecs, so the formats list reflects what your installation can actually produce.

Changed

• FFmpeg libraries are now verified with cosign signatures for added security.

Changed

• Cleaned up and reorganized deployment configuration files

Added

• Linux tarball release artifact for installing zshot on non-Debian/Ubuntu distributions, including an INSTALL.txt guide and a list of required runtime dependencies

Changed

• Docker runtime image switched to a slimmer base, reducing image size
• Release artifacts are now accompanied by an SBOM and scanned for known CVEs before publishing

Added

• LLM-powered navigation now supports web search tools when the model requests them (Anthropic and OpenAI providers)
• New READ action available to the AI navigator, allowing it to read page content during a session
• AI navigator can now specify custom wait conditions between steps rather than using a fixed delay
• Screenshots sent to vision models can now be annotated with element overlays to improve navigation accuracy

Changed

• Model connection strings are easier to specify — shorter, more ergonomic syntax for pointing at local or remote LLM providers

Fixed

• Corrected protocol scheme strings used in LLM output parsing

Added

• stdio navigator: a new --navigator mode that lets an external agent drive browser navigation by reading/writing over standard I/O, enabling fully scriptable or AI-controlled capture workflows
• Navigator session state can now be passed through to the formatter, allowing context gathered during navigation to influence output formatting
• New AGENT_HELP.md and skill reference documentation for agent-based navigation

Changed

• Update CEF

Changed

• Increased timeouts for LLM output types to better accommodate slower model response times

Added

• Built-in ad/tracker blocking: pass one or more filter-list files (in standard Adblock Plus / uBlock Origin syntax) to block network requests while capturing

Changed

• Clarified help text for the ad-blocking filter-list option

Added

• IP blocklist support: pass a blocklist file to prevent the browser from making requests to specific IP addresses or CIDR ranges during capture. Useful for restricting outbound connections in sensitive or sandboxed environments.

Fixed

• Docker images now always include up-to-date CA root certificates.

Fixed

• Docker images now include essential administrative utilities
• Artifact signing now correctly handles nested directory structures

Added

• New --chromium-arg flag to pass arbitrary Chromium arguments directly to the browser
• New container mode for running in containerized environments
• arm64 (Linux aarch64) builds are now available

Fixed

• Resolved a conflict with Homebrew-installed FFmpeg that could cause incorrect library detection

Added

• Improved accessibility tree (axtree) representation used by the AI navigator, aligned with modern agent browser conventions for more accurate element identification and interaction
• Support for iframes in axtree snapshots

Changed

• Navigator action parsing and element selection logic updated to work with the revised axtree format, improving reliability when the AI agent interacts with page elements

Added

• Invalid flag combinations are now caught early and reported as clear errors, rather than producing unexpected output or crashing

Changed

• Update CEF

Added

• Full-page capture now explicitly rejects animated formats (GIF, WebM, MP4) and video output, preventing silent failures or unexpected results

Changed

• Full-page capture errors are now reported as clear messages rather than unexpected crashes

Fixed

• Various edge-case panics replaced with proper error handling across capture and rendering

Added

• Direct integration with OpenAI and Anthropic APIs, replacing the previous library-based approach — improves reliability and avoids deprecated behaviour

Changed

• Updated AI/LLM provider internals to communicate directly with OpenAI and Anthropic; connection strings and provider configuration remain compatible
• Expanded skill documentation covering accessibility and performance capabilities

Changed

• Improved reliability of LLM output parsing, with better handling of edge-case JSON responses and more appropriate default token limits for Anthropic models

Changed

• Support contact updated to link to zshot support page

Added

• --eula flag to view the end-user license agreement

Changed

• Update CEF to latest version

Fixed

• Improved full-page capture to correctly handle pages with inner scrollable elements and fixed sidebars

Fixed

• Fixed typos in startup/initialization messages and added pointers to the canary build for pre-release testing

Added

• Performance tracing output — capture a Chrome-style JSON trace file alongside your screenshots/videos using the new --output-type trace option, useful for diagnosing page load performance
• pprof profiling output — capture CPU profiling data in pprof format (--output-type pprof) for deeper performance analysis
• Dark mode emulation — force pages into dark color scheme via prefers-color-scheme: dark using the new --dark-mode flag
• Accept-Language header emulation — send a custom Accept-Language header with requests to control locale-aware content
• Timezone emulation — pair timezone settings with geolocation for accurate locale emulation
• GPU rendering control — GPU hardware acceleration is now the default; a --software-rendering flag is available to fall back to software rendering
• Autoplay support — media on captured pages will autoplay where possible; audio is always muted to the host OS
• WARC payload digest — WARC output now includes a correct payload digest field
• GFM strikethrough portability — Markdown output now renders strikethrough in a more portable way across GFM renderers
• HTML output base URL — HTML output now includes a <base href> tag for improved portability of captured pages
• --wait-for timer mode — timer-based waits now reflect true elapsed time rather than waiting for the page load event

Changed

• Default background is now white — pages with a transparent or unset background now render with a white background instead of black
• GPU rendering enabled by default on Linux — broader GPU feature set is now unlocked for improved rendering fidelity
• --full-page validated against output type — passing --full-page for output types where it is not applicable (e.g. video, PDF) now produces a clear error instead of being silently ignored
• Unknown request fields rejected — the HTTP server now returns an error for unrecognised JSON fields in capture requests, preventing silent misconfiguration
• LLM JSON prompting improved — prompt templates now include additional guidance to help smaller models produce well-formed JSON output reliably
• JPEG EXIF metadata — the software metadata field is now written to the correct EXIF tag for JPEG output
• Touch point limit enforced — device emulation now caps touch points to the maximum supported by the browser engine, with a clear error if the requested device exceeds it
• robots.txt fetches now subject to IP filtering — the same SSRF/IP-filter rules applied to page requests are now also enforced when fetching robots.txt
• Improved --help output — help text reorganised and expanded, including guidance for agent/automation use

Removed

• Animated WebP output temporarily removed — animated WebP captures are not available in this release pending a fix; static WebP output is unaffected

Changed

• Update stale licenses

Added

• Updated FFmpeg instructions for binary relinking

Changed

• Updated JSON schema with additional fields

Changed

• Binary downloads now use public URLs instead of direct storage URLs

Added

• Skills can now be installed directly into a zip file

Added

• Output can now be written to stdout by passing - as the output path, enabling piping to other tools
• New skill install subcommand to wire zshot into an AI agent as a tool

Changed

• PNG and PDF output formats now require a Standard license or higher; HTML and MHTML are available on the free tier

Fixed

• Restored missing signing key for v0.1.1 release verification

Added

• New --request-schema flag for consuming the request schema in other apps

Added

• --image-selector and --video-selector: capture only the contents of a specific element on the page (for image and video output respectively)
• --wait-for-css: wait for a specific CSS selector to appear in the DOM before capturing

Fixed

• Fix Google model names being prefixed w/ models/

Added

• Support for Groq as an LLM provider
• Support for listing available models when using the Google LLM provider
• Ollama connections can now use HTTPS (e.g. for remote/proxied Ollama instances)

Added

• llm-info command now validates your LLM configuration and reports any issues with your setup

Added

• New --llm-info flag to print details about available LLM models and their capabilities
• New --llm-info-schema flag to print the JSON schema for LLM info output
• When using AI-powered navigation or formatting from the CLI, the local model is now automatically selected as the default — no need to specify it explicitly
• Improved config file lookup: zshot now searches a broader hierarchy of locations to find your configuration file

Changed

• Video capture now more reliably meets the requested frame rate, reducing dropped frames during recording
• Update CEF

Changed

• Removed stale references to Apple/macOS-specific options in CLI help text — argument descriptions are now platform-neutral

Added

• Device listings now include the device ID, making it easier to reference specific devices in other commands

Fixed

• Improved handling of symlinked installations so the CLI resolves paths correctly when run via a symlink

Fixed

• Fixed a path handling bug on Windows where forward slashes in CEF download paths were not converted correctly.
• Improved error messages when file operations fail during CEF setup (errors now include the affected file path).

Added

• MCP server now fetches and exposes only the LLM resources available to the connected client’s tooling, rather than a static full list

Added

• The MCP server now advertises screenshot resources, making them discoverable by MCP clients

Changed

• The MCP server now always uses the zshot:// URI scheme for resources, replacing the previous context-dependent scheme selection

Changed

• MCP tool responses now always return structured content wrapped in a consistent envelope, regardless of output type

Added

• MCP clients now receive proper error responses when a page fails to load or a timeout occurs, instead of hanging or receiving a generic failure

Fixed

• Load errors and timeout errors are now correctly propagated to MCP tool calls

Added

• MCP: resource link handoff — when a captured file is too large to embed directly in the MCP response, the server now returns a resource link instead, so callers can retrieve the file separately rather than hitting framing limits.

Changed

• MCP: HAR capture no longer records request/response bodies by default, reducing output size for typical capture workflows. Body capture can still be enabled explicitly.
• MCP: tool and resource descriptions updated for clarity.

Changed

• Improved MCP server tool descriptions for clearer configuration guidance

Added

• License keys are now forward-compatible: older license key blobs can be used on newer versions of zshot without needing to regenerate keys on every update.

Added

• Initial version