Skip to content

changelog.mdx

Latest commit

 

History

History
506 lines (407 loc) · 50.8 KB

changelog.mdx

File metadata and controls

506 lines (407 loc) · 50.8 KB
Raw
title Changelog
description Updates and improvements

import { YouTubeVideo } from '/snippets/youtube-video.mdx';

For API library updates, see the [Node SDK](https://github.com/onkernel/kernel-node-sdk/blob/main/CHANGELOG.md), [Python SDK](https://github.com/onkernel/kernel-python-sdk/blob/next/CHANGELOG.md), and [Go SDK](https://github.com/onkernel/kernel-go-sdk/blob/main/CHANGELOG.md) changelogs.

<Update label="July 31" tags={["Product", "Docs"]}>

Product updates

  • Released @onkernel/eve-extension v0.1.3, a Kernel-powered browser extension for Vercel's Eve agent.
  • Published a Codex plugin packaging for the Kernel skills repo, so Codex users can install the full Kernel skill set the same way Claude Code and Cursor users can.
  • Added a get_telemetry action to the MCP server's manage_browsers tool for reading archived browser telemetry — including for deleted sessions and events captured before telemetry was turned off — with category filters, time windows, and pagination.
  • Expanded managed auth: browser telemetry is now configurable per connection and captured on timeline events, the health-check interval is adjustable from the dashboard, and health checks and their replays now appear on the connection timeline alongside logins and re-auths.
  • Extended @onkernel/cua-agent with Claude Opus 5 and Moonshot Kimi K3 computer-use providers, semantic browser waits (waitFor conditions instead of fixed sleeps), verified browser action plans, and gated Anthropic's native computer_20260701 / browser_20260701 tools by model.
  • Added named markers to browser replay recordings, exposed as MP4 chapters so playback jumps directly to key moments.
  • Extended the CLI browser-pools commands with telemetry configuration, matching browsers create/update.
  • Improved kernel upgrade in the CLI to refresh the Homebrew tap before upgrading to pick up latest version, and made CLI commands fail fast on interactive prompts in non-interactive shells to avoid hanging scripts.
  • Added a complete audit log download helper to the Go, Node, and Python SDKs.
  • Added support for solving PerimeterX / HUMAN "Press & Hold" challenges in stealth-mode browsers.
  • Added customer-supplied CA bundles to custom proxies for BYO TLS-terminating (MITM) proxies.
  • Improved kernel audit-logs search output to show user IDs as a separate column.
  • Shipped inline text editing and document bookmarks (with a public bookmarks agent API) in just-html.

Documentation updates

<Update label="July 17" tags={["Product", "Docs"]}>

Product updates

  • Launched browser telemetry: capture events from inside a browser session — console output, network activity, page lifecycle, user interactions, captcha solves, and operational signals like crashes — then stream them live or pull them later for analysis, with opt-in control over which categories you capture. Alongside the launch, we added an in-VM OTLP export sink and OTel-compatible support in the browser events API, so telemetry can be shipped to any OpenTelemetry-compatible collector.
  • Added new Grok 4.5 and Meta Muse Spark computer-use providers to @onkernel/cua-agent, along with a --proxy flag for routing the agent's browser through a Kernel proxy, configurable retries on transient provider errors, and opt-in recovery when a provider returns an exact-empty response.
  • Gave kernel audit-logs in the CLI a download subcommand for chunked bulk exports and the ability to exclude multiple HTTP methods from search, so audit workflows can pull large windows without hitting pagination caps.
  • Improved cold-start performance: reduced Chromium restart time, and cut ~5s off session creation for profile-restore paths by fixing a chromium-launcher port check that blocked on stale sockets.
  • Set the default fill-rate for new browser pools created from the dashboard or CLI to 25%, so pools warm up more predictably out of the box.
  • refresh_on_profile_update now defaults to true when a browser pool has a profile attached, and auto-unsets when the profile is removed, so pooled sessions stay in sync with the underlying profile without manual configuration.
  • Expanded managed auth: the dashboard can now manage credential fields and TOTP secrets directly, and reauth now selects an already signed-in account on the SSO account chooser, so re-authentications can complete without human intervention when the browser is already signed in.
  • Shipped a Cmd+K command palette to the dashboard for jump-to-anywhere navigation.
  • Published hermes-browser-plugin, a new Kernel cloud browser provider plugin for Hermes Agent.
  • Improved just-html with section deeplinks, comment permalinks, and a public integration-discovery metadata endpoint, so shared docs are easier to navigate, link, and index by agents.

Documentation updates

  • Published a new Zero Data Retention page covering Kernel's ZDR posture.
  • Documented a "managed proxy without CAPTCHA solver" pattern for stealth BYO proxy setups.
  • Moved the scaling guide into Introduction as a dedicated Scale page.
  • Documented in the browser telemetry guide which categories are disabled under BAA.

<Update label="July 10" tags={["Product", "Docs"]}>

Product updates

  • Consolidated on-demand and browser-pool concurrency into a single unified concurrency limit that counts browsers.create() sessions and pool reservations against one org cap. max_pooled_sessions is deprecated; max_concurrent_sessions now covers both. Rolling out gradually.
  • Shipped kernel audit-logs search in the CLI for querying the org audit log by time window, HTTP method, service, auth strategy, and user, with table or JSON output. GET requests are excluded by default to keep the signal on mutations.
  • Added a refresh_on_profile_update flag to browser pools (SDK, CLI --refresh-on-profile-update, and dashboard). Opted-in pools automatically flush their idle browsers when the pool's managed-auth profile re-authenticates, so pooled sessions stay logged in after a re-auth.
  • Extended @onkernel/cua-agent with action-plane modes — computer, browser, and hybrid — plus setMode() and a /mode runtime switcher, and wired in support for Anthropic's shipped native computer_20260701 and browser_20260701 tools.
  • Exposed profile_id and extension_ids on browser pool reads, and returned an SHA-256 checksum for uploaded extensions.
  • Added the ability to inspect credential value keys and remove specific values on managed-auth credential updates.
  • DELETE /org/api_keys/{id} now returns 400 cannot_delete_current_key when the target key is the one authenticating the request, so scripts and Terraform-style provisioners can't brick their own credentials mid-run.

Documentation updates

<Update label="July 3" tags={["Product", "Docs"]}>

Product updates

  • Shipped kernel extensions get <id-or-name> in the CLI for extension metadata lookups, with table or JSON output.
  • Extended CLI browser telemetry with kernel browsers telemetry events <id> for paginated historical reads and a --replay=all flag on browsers telemetry stream to replay from the oldest retained event.
  • Simplified browser pool profiles to id/name-only, dropping the --save-changes flag on browser-pools create/update; single-session browsers create --save-changes is unaffected.
  • Extended @onkernel/cua-ai with computer-use support for the claude-sonnet-5 model.
  • Threaded previous_response_id through the Tzafon and OpenAI computer-use providers in @onkernel/cua-ai, so multi-turn runs send only the latest screenshot delta instead of replaying the full history each turn.
  • Rebuilt the dashboard managed-auth connection detail page around an overview/history/config tab layout, with a pinned stats row and a full event timeline (logins, re-auths, health checks) filterable by type.
  • Brought mobile proxy creation into the dashboard, with country, state, and city targeting.
  • Expanded the dashboard pool create/edit dialogs with tablet (768x1024) and mobile (390x844) viewport presets.

Documentation updates

  • Published a Computer Use overview page covering the shared screenshot-predict-execute loop and a guide for building custom agents with @onkernel/cua-agent.
  • Clarified that browsers released back into a pool with the default reuse: true keep their original configuration, and that browserPools.update() only rebuilds idle browsers when discard_all_idle: true is passed.

<Update label="June 26" tags={["Product", "Docs"]}>

Product updates

  • Added a playwright_execute tool to @onkernel/cua-agent for combining computer use with Playwright automation, plus support for gemini-3.5-flash as a CUA model.
  • Extended the MCP server's browser pool tool with SDK parity fields (start_url, chrome_policy, kiosk_mode, profile and viewport settings, fill_rate_per_minute) and a structured acquire/release workflow. Added parity for actions across manage_apps (delete_deployment, app/deployment search), manage_proxies (get, check with optional check_url), manage_profiles (get, paginated search), and manage_projects (get_limits, update_limits).
  • Added a start_url parameter to browser pool acquire, letting each acquired session override the pool's default start URL.
  • Added --chrome-policy flags to browsers create/update and browser-pools create/update in the CLI for setting Chrome popup, PDF, and homepage behavior from the terminal.
  • Extended kernel browsers update in the CLI with --name and --tag parameters, so a running session's name and tags can be changed after creation alongside the existing browsers create/acquire flags.
  • Brought custom session names and tags into the dashboard: the launch dialog now accepts an initial set of tags, sessions can be renamed and re-tagged after creation, and tag chips on the sessions list are click-to-filter — so the names and tags set via the API are actionable from the UI, making it easier to find a specific session out of many concurrent ones.
  • Added an API key rotate endpoint (POST /org/api_keys/{id}/rotate) that issues a replacement key, copies the rotated key's name and project scope, and keeps the old key working for a configurable grace period (default 7 days).
  • Shipped a generate-video skill for producing smooth, deterministic MP4 videos from a web scene — used internally for launch and explainer clips.

Documentation updates

  • Enabled dark mode on the docs.
  • Added a common use cases section to the policy.json page with browsers.create() examples for popup blocking, PDF inline rendering, and homepage overrides.
  • Documented X-Kernel-Project-Id header resolution rules and direct project and extension lookups on the Projects page.
  • Documented the API key rotate endpoint, including the optional days_to_expire and expire_in_days parameters.

<Update label="June 12" tags={["Product", "Docs"]}>

Product updates

  • Released the Kernel × Claude Managed Agents cookbook, a public recipe for running parallel computer-use agent swarms with Claude Managed Agents on Kernel cloud browsers, with API keys secured via Vaults.
  • Shipped a kernel-browser-harness skill for setting up Kernel browser sessions inside Claude Code with persistent session IDs across calls.
  • Extended kernel browsers acquire in the CLI with name and tags parameters, so pooled browsers can be labeled at acquisition time alongside browsers create.
  • Made browser telemetry categories opt-in. The --telemetry flag (and telemetry request config) now records only the categories you list; enabling telemetry without specifying categories ships a small default set instead of every category.

Documentation updates

<Update label="June 5" tags={["Product", "Docs"]}>

Product updates

  • Browser sessions now play audio by default, and the browser replay recording API gained a record_audio option to capture that audio in replay videos.
  • Extended browser telemetry to the CLI: use --telemetry=all|off|<list> on browsers create and browsers update to configure telemetry per session, and kernel browsers telemetry stream <id> to tail live events with category and type filters and SSE resume support.
  • Added API key management commands to the CLI for creating, listing, retrieving, updating, and deleting org API keys directly from the terminal.
  • Expanded the MCP server from 10 to 16 tools, adding managed auth (manage_auth_connections, manage_credentials, manage_credential_providers), browser_curl for HTTP(S) through a live browser, API key management, and project management. MCP-created browser sessions also now accept the same configuration as the API.
  • Browser sessions can now be given a custom name at create time, making them easier to identify in the dashboard and audit logs.
  • Improved kernel status to distinguish between the API being down versus unreachable.

Documentation updates

  • Reorganized the MCP server documentation into per-page navigation for easier browsing.
  • Expanded and regrouped the CLI reference to match the MCP docs structure.
  • Added Go SDK code examples across the docs.
  • Added a managed auth CLI reference page.
  • Clarified that ISP proxies provide a static IP across sessions.

<Update label="May 28" tags={["Product", "Docs"]}>

Product updates

  • Soft-launched auth.md, a discovery file that lets agents sign up for a Kernel account or fetch an API key for an existing user account on their own. Point your agent at the URL and it can self-provision. Based on the new auth.md protocol from WorkOS.
  • Added browser telemetry as a first-class request config. Pass telemetry.enabled (and per-category toggles) on browsers.create() or browsers.update() to capture telemetry for the session, and stream it live from GET /browsers/{id}/telemetry/stream (SSE).
  • Exposed API key management in the Node, Python, and Go SDKs. Create, list, retrieve, update, and delete keys on /org/api_keys programmatically.
  • Promoted can_reauth_reason on ManagedAuth to a typed enum (14 documented values like requires_totp_without_secret, no_viable_plans, requires_external_action) so SDK consumers can branch on it directly instead of string comparisons.
  • Added an Auto Re-Auth / Needs Human capability chip to each row on the dashboard /auth page, with a tooltip mapping each can_reauth_reason to a human-readable explanation.
  • Added TOTP secret key support to managed auth credentials, so one-time passwords are generated automatically during login and re-authentication. No human intervention required.
  • Updated the live view loading screen to a progress bar for clearer visual feedback during browser startup.
  • The /projects/* endpoints are now routed under /org/projects/*. The previous paths are deprecated.

Documentation updates

  • Added new Create, Control, and Observe introductory guides covering the core primitives for getting started with browser automation.
  • Expanded the Control guide with three patterns for combining computer use with playwright.execute: expose playwright.execute as a tool, checkpoint state between CUA steps, and drop the agent once selectors stabilize for replay at scale.
  • Documented the new health_checks and auto_reauth connection flags on managed-auth connections, including defaults, how they interact (auto_reauth only applies while health_checks is on), and a create-time example with both disabled.
  • Updated the MCP server tool reference to match the current API.
  • Clarified that stealth mode's default proxy is ISP, not residential. ISP proxies use residential ASNs on datacenter infrastructure.

<Update label="May 21" tags={["Product", "Docs"]}>

Product updates

  • Added auto standby to browser pool instances — pools can now automatically suspend when idle and resume on the next incoming request, reducing costs without manual intervention.
  • Launched new browser configuration options on browsers.create() and browser pool definitions: chrome_policy and start_url for opening directly to a specified URL on launch. The corresponding --start-url flag is available in the CLI.
  • Exposed full Projects CRUD in the public API, so projects can be created, updated, and deleted programmatically alongside the existing list endpoint.
  • Managed auth improvements: Added health check and automatic re-authentication controls to the managed auth API, letting you configure check intervals and reauth triggers per connection. Sessions are now automatically recorded — when a connection enters NEEDS_AUTH on the dashboard, a "View last login attempt" link routes directly to the session replay, and /auth rows now show a re-auth capability chip. Also, broader SSO and OAuth provider coverage, support for Google's two-step mobile prompt flow, per-connection post-login wait configuration (post_login.wait_ms), automatic SSO provider brand icons in auth dialogs, MFA alternatives displayed on the external action waiting screen, and a post-login browser refresh before the profile snapshot is captured for cleaner saved sessions.
  • Extended proxy.check() with an optional url parameter for testing reachability against a specific target domain before assigning the proxy. This is useful for catching proxies that pass generic health checks but are blocked on your target site.
  • End of life'd persistent browsers. If you were using persistence, we suggest timeout_seconds with Profiles.

Documentation updates

  • Added a new hCaptcha page documenting beta support for hCaptcha solving.
  • Refreshed managed auth documentation for May 2026: new dedicated connection lifecycle page covering health checks and re-authentication, a shared connection configuration reference, documented success_url / error_url query parameters for the hosted UI, start_url references across browser and pool docs, a reorganized sidebar, and new FAQ entries for short-session reauth and multi-step login forms.
  • Clarified that managed residential proxy IPs are stable within a session but are not guaranteed to persist across sessions.
  • Updated the Yutori integration guide to Navigator n1.5.
  • Clarified that profile updates do not propagate to idle browser pool instances — pools must be recycled for profile changes to take effect.

<Update label="May 8" tags={["Product", "Docs"]}>

Product updates

  • Released @onkernel/managed-auth-react, a drop-in React component library for embedding managed auth flows directly into your app. Ship a Kernel-powered login experience without rebuilding the credential entry, MFA, and SSO dialogs yourself.
  • Added a Docker Sandboxes mixin kit for running Kernel inside Docker's AI sandboxes (sbx). The kit ships the Kernel CLI, Claude Code skills, and a proxy-managed auth header pre-configured, so agents inside the sandbox can call the Kernel API while your real KERNEL_API_KEY stays on the host.
  • Extended Projects: profile, browser pool, extension, and credential names are now scoped per-project, and the GET /projects API and dashboard project selector support server-side search and pagination.
  • Made the 1Password credential dropdown searchable in managed auth dialogs.
  • Improved managed auth autofill reliability on multi-step sign-in pages.
  • Added -o json output to kernel browsers playwright execute in the CLI, matching the other browsers subcommands, so script runs can be piped into other tooling.

Documentation updates

  • Documented in the computer controls docs that computer controls screenshots are faster than CDP screenshots on WebGL-heavy pages.
  • Updated kernel-cli skill examples in the Kernel skills repo and documented the return value of playwright execute so agents can use the result in follow-on steps.

<Update label="May 1" tags={["Product", "Docs"]}>

Product updates

  • Added kernel browsers curl command to the CLI for making HTTP requests directly from a browser session using browser's cookies, session state, and network identity.
  • Extended Projects support to the CLI and MCP server — use the --project flag to scope CLI commands to a specific project, and MCP tools now resolve projects by name.
  • Improved mouse movement timing with Gaussian-distributed delays between movements for more natural-looking interactions.
  • Improved managed auth UX: MFA prompts now display which channel (email or phone) the verification code was sent to, and --credential-auto now defaults to enabled when --credential-provider is set.
  • Added a dedicated switch MFA option type so generic method-switcher links ("Use another method", "Try another way") are distinguished from buttons that name a specific method, improving auto-click reliability on multi-step MFA pages.
  • Managed auth CUA flow now resolves external credentials (e.g. 1Password) via auto-lookup.
  • Managed auth connections now expose browser_session_id so callers can inspect or terminate the underlying browser session via the /browsers API.
  • Added a diff-profile-archives skill to the Kernel skills repo for comparing Chrome profile archives directly with a coding agent.

Documentation updates

<Update label="April 24" tags={["Product", "Docs"]}>

Product updates

  • Launched Projects so you can keep production and staging separate, split resources between teams, scope API keys to a single environment, and cap concurrency per project so one workload can't exhaust your org quota. Existing resources have been moved into a Default project, so nothing changes for current workloads.
  • Kernel is now a Cloudflare Web Bot Auth partner. Kernel browsers can cryptographically identify themselves to Cloudflare-protected sites, so Kernel traffic isn't blocked.
  • Anti-detection features are now on by default for all Kernel browsers. Stealth mode now specifically adds the managed ISP proxy and CAPTCHA solver (both opt-out), and non-stealth browsers fully support custom proxies.
  • Revamped the managed auth hosted login page: all available sign-in options (password fields, SSO providers, MFA, alternate sign-in methods) now render together on a single page, so end users can pick the path they want, instead of being funneled through one at a time.
  • Managed auth input fields now display contextual helper text when the site surfaces hints, reducing user confusion on multi-step logins.
  • The "Save credentials after login" option is now automatically disabled when 1Password is selected as the credential source, since those credentials are already managed externally.
  • Kernel now supports WebSocket connections through its API, enabling process attach and other long-lived streaming workflows.
  • Exceeding invocation concurrency limits now returns a 429 response immediately, for faster and more actionable feedback.

Documentation updates

<Update label="April 17" tags={["Product", "Docs"]}>

Product updates

  • Managed auth now scrolls to find and click submit buttons that are below the initial viewport.
  • Added a Tzafon Northstar computer use template to the CLI for building Northstar-based browser agents.
  • Browser search now matches by pool name in addition to session ID, profile, and proxy.
  • Added an enterprise option to remove the Kernel logo from the live view loading screen.
  • Eliminated live view scrollbar flickering on mobile viewports.
  • Fixed Safari clipboard paste issues in live view iframes.

Documentation updates

  • Added a new Tzafon Northstar integration guide.
  • Updated the bot anti-detection page to document smooth Bezier curve mouse movements, and renamed the Chrome policy configuration page in the nav for clarity.
  • Added documentation for smooth typing with typo injection — new smooth and typo_chance parameters for human-like typeText behavior.
  • Updated DPA subprocessor list URLs to point to the new trust center at trust.kernel.sh.

<Update label="April 10" tags={["Product", "Docs"]}>

Product updates

  • Improved live view scroll feel for smoother, more natural scrolling.
  • We removed idle disk space charges for browser pools on the startup tier too, so you'll no longer be billed for storage when your pools aren't in use.
  • Improved browser creation speed when loading profiles by reducing the CDP ready wait to near-instant.
  • Added human-like smooth typing to cloud browsers, with optional typo injection for more natural-looking interactions.
  • Launched a Cursor Marketplace plugin bundling all Kernel skills, an MCP server for cloud browser management, and best practices.

Documentation updates

  • Documented MFA token auto-retry behavior for managed auth sessions.
  • Added a new policy.json page to the Reserved Browsers documentation.
  • Clarified that profiles can have multiple auth connections.
  • Added a Headful + GPU acceleration option to the pricing calculator.

<Update label="April 3" tags={["Product", "Docs"]}>

Product updates

  • Added spending cap management to the dashboard billing page, allowing org admins to set and update spending limits directly from the UI. Currently in limited preview.
  • Enhanced browser stealth with additional evasion techniques.
  • Fixed navigator fingerprint gaps (maxTouchPoints, userAgentData brands, screen.availHeight) to better match real desktop Chrome behavior.
  • Added disable_default_proxy option for stealth browsers, allowing users to disable the default proxy on stealth-mode browsers.
  • Improved error responses for browser process start failures.

Documentation updates

<Update label="March 27" tags={["Product", "Docs"]}>

Product updates

  • Added support for custom Chrome policy.json on browser pools, enabling users to configure Chrome settings like popups, PDF behavior, and homepage without manual workarounds.
  • Updated the CLI to prioritize API key authentication over OAuth, and added new commands and flags.
  • Improved browser pool acquisition speed.
  • Improved CDP proxy reconnect behavior after Chromium restart, preventing broken browser sessions.
  • Corrected Safari clipboard sync in the dashboard live view, resolving copy/paste issues for users viewing browser sessions from Safari.
  • Improved mouse movement in live view.

Documentation updates

<Update label="March 20" tags={["Product", "Docs"]}>

Product updates

  • Launched GPU acceleration for cloud browsers. GPU acceleration attaches a virtual GPU to your cloud browser, dramatically speeding up agents on sites that use canvases, are heavy on visualizations, or rely on WebGL. In our internal benchmarks, GPU-accelerated browsers hit 60 fps —6x faster than regular browsers— making browsing feel faster pretty much everywhere on the internet. GPU acceleration is available as a research preview on Startup and Enterprise plans.
  • Increased Chromium's default maximum connections per proxy to 16, improving throughput for proxy-based browser sessions.
  • Updated the live view logo and favicon to reflect the latest Kernel branding.
  • Updated OpenAI CUA templates to use Kernel's native Computer Controls API instead of Playwright for improved reliability and performance.
  • Fixed an issue where deleted browser sessions couldn't be listed or downloaded in session replays via the CLI.

Documentation updates

<Update label="March 13" tags={["Product", "Docs"]}>

Product updates

  • Added WebDriver BiDi support for cloud browsers through a new partnership with Vibium. AI agents can now use Vibium to connect to Kernel and navigate pages, fill forms, click buttons, and take screenshots across thousands of cloud browsers running in parallel.
  • Launched Kernel Eval Protocol, an open-source repo that pairs Kernel Browser Pools with Fireworks RFT and Eval Protocol to evaluate and fine-tune VLM browser agents using reinforcement learning.
  • Added a usage field to the browser session details API response, which returns session metrics like uptime_ms. Use this to track session duration and build billing or analytics workflows around your browser usage.
  • Extended smooth Bezier curve mouse movements to click and drag operations, and added --smooth and --duration-ms flags to the move-mouse command in the CLI for more natural-looking browser interactions.
  • Improved headless viewport resize performance by using a CDP fast path, reducing resize time from ~5 seconds to ~10 milliseconds.
  • Added support for runtime proxy configuration swapping, enabling you to change a browser's proxy settings mid-session without restarting.

Documentation updates

<Update label="March 6" tags={["Product", "Docs"]}>

Product updates

  • Added support for mobile and tablet viewports, enabling browser automation at phone and tablet screen sizes.
  • Added a kernel status command to the CLI for checking API and service health at a glance.
  • Added a --force flag to kernel browsers update for resizing the viewport during an active recording, which gracefully stops and restarts the recording.
  • Improved Managed Auth MFA handling by resolving MFA options by label, type, or display string for more reliable multi-factor authentication flows.
  • Enhanced auth connection output in the CLI with richer details from kernel auth connections get and kernel auth connections list.
  • Added a Pool column and --query flag to kernel browsers list in the CLI for easier filtering and identification of pooled browsers.
  • Updated the Anthropic computer use template to default to use claude-sonnet-4-6 for improved agent performance.
  • Fixed start session replay error.
  • Updated localStorage to persist across profile sessions.
  • Fixed scroll notch count handling in computer use templates.
  • Fixed an anti-echo guard issue that could cause video sync problems in live view.

Documentation updates

<Update label="February 27" tags={["Product", "Docs"]}>

Product updates

  • We launched our new look! Our visual identity has been updated across our website, product, and docs.
  • Important: Persistent browser functionality will be end-of-life'd as of May 18th. If you're using this functionality, you've received an email from us with next steps. Please reach out for further support.
  • Improved the MCP server by adding 21 new tools covering browser pools, computer controls, proxies, extensions, and more. Consolidated 36 individual tools into 10 streamlined tools and enhanced the computer_action tool to support executing multiple actions in a single call.
  • Added API support for clipboard operations, enabling programmatic copy/paste between your automation code and the remote browser.
  • Introduced human-like Bezier curve mouse movements for more natural-looking browser interactions that better evade bot detection.
  • Added an option to disable live view logging, reducing noise when debugging browser sessions.
  • Added kernel deploy delete and kernel app delete commands to the CLI for removing deployments and apps.
  • Added server-side session ID search to the dashboard, allowing users to find past browser sessions by ID without loading them all locally.
  • Added profile pagination and search to the dashboard and CLI, fixing crashes for users with large numbers of profiles.
  • Pre-installed ssh and websocat in browser images for easier debugging and WebSocket testing.
  • Added more fonts to browser images for better rendering of international content.
  • Improved profile upload performance by buffering zstd archives in memory instead of writing to temp files.

Documentation updates

  • Added a new Yutori computer use integration guide with setup instructions and code examples.
  • Clarified viewport configuration recommendations with guidance on choosing the right resolution for your use case.

<Update label="February 20" tags={["Product", "Docs"]}>

Product updates

  • Added mouse position tracking to the CLI, enabling retrieval of current mouse coordinates.
  • Updated Yutori computer use templates to support the n1-latest model for improved agent performance.
  • Updated Managed Auth by adding subdomain-based sign-in support, better error handling for 401 Unauthorized and 410 Gone responses, and enhanced error messaging with structured error codes and actionable guidance.
  • Improved browser display by auto-toggling Chromium app mode on small viewports for a cleaner, more immersive experience.
  • Fixed screen resize accuracy by removing unnecessary rounding in ChangeScreenSize to ensure pixel-perfect display dimensions.

Documentation updates

  • Enhanced secrets documentation with practical examples for LLM-powered applications and detailed guidance for deploying apps with environment file configurations.

<Update label="February 13" tags={["Product", "Docs"]}>

Product updates

- Launched [Web Bot Auth](/browsers/bot-detection/web-bot-auth) in partnership with Vercel, enabling agents to cryptographically sign requests and prove they're legitimate instead of getting blocked by bot detection. - Released [Managed Auth](/auth/overview), simplifying authentication by securely logging into any site without custom auth flows or exposing credentials to the LLM, and maintaining up-to-date credentials. - Added a `POST /computer/batch` [endpoint](https://kernel.sh/docs/api-reference/browsers/execute-a-batch-of-computer-actions-sequentially) for executing multiple computer actions in a single API call, reducing round-trip latency for complex automations. - Improved the [CLI](https://github.com/onkernel/cli) by adding new commands for managing auth connections, supporting `-o json` output for `kernel ssh --setup-only`, allowing pool names as positional arguments in `kernel browser-pools create`, and enabling file exclusions when publishing extensions. - Improved input reliability with context-aware timing in key press and mouse drag operations. - Fixed an issue where stealth mode browsers couldn't access plain HTTP websites due to forced HTTPS redirects.

Documentation updates

  • Clarified pricing for headful browser sessions.
  • Added comprehensive Managed Auth documentation, including billing guidance.

<Update label="February 6" tags={["Product", "Docs"]}>

Product updates

  • Added a /json endpoint to the CDP proxy, enabling native Playwright connectOverCDP() connections.
  • Launched an API to get the current cursor position, which is useful for debugging and building adaptive automation logic.
  • Released an API to retrieve browser session IDs associated with a specific invocation.
  • Added SSH access to Kernel browsers for debugging, running commands, and tunneling your local development server to the browser.
  • Enabled support for the 1280x800 viewport size, which improves click coordinate accuracy.
  • The Live View clipboard now syncs correctly between browser and local machine.
  • Chrome address autofill prompts are now automatically disabled.

Documentation updates

  • Added a pricing calculator to help you estimate costs based on your plan, session length, and browser type.
  • Clarified file access limitations and browser session deletion behavior.

<Update label="January 30" tags={["Product", "Docs"]}>

Product updates

  • We changed the default search engine for Kernel browsers to DuckDuckGo, which reduces CAPTCHA friction during web searches.
  • We updated the Gemini Computer Use Python and TypeScript templates to use Kernel's Computer Controls API.
  • We released skills for Kernel, so you can easily build browser agents directly from Claude Code.

Documentation updates

  • We published a Scaling in Production guide with recommendations and best practices for handling high-concurrency workloads. Read it to learn how and when to use browser pools versus on-demand browsers, and see code samples for production implementations.
  • Added a new integration guide for Agent Browser, Vercel's headless browser automation CLI, which has built-in support for Kernel browsers.

<Update label="January 23" tags={["Product"]}>

Product updates

- [Browser pools](/browsers/pools/overview) are now generally available for our Startup and Enterprise plans! Make your startup times even faster with a fleet of instant browsers that come pre-loaded with the logins, cookies, and extensions your agents depend on. - We improved the Browsers page user experience. Enjoy instant clickability of browser cards, shift+click range selection, preserved view state when switching between grid and list views, a live toggle for auto-refresh, keyboard shortcut support (Cmd+B / Ctrl+B to toggle sidebar), and more. - We added support for viewing [replays](/browsers/replays) of deleted browser sessions. Deleted browsers now navigate to their detail page, which shows the replay player with multiple replays displayed as tabs. The API now supports fetching deleted browser details using new query parameters. - Overhauled our [Anthropic Computer Use template](https://github.com/onkernel/cli/tree/main/pkg/templates/typescript/anthropic-computer-use) to use Kernel's [Computer Controls](/browsers/computer-controls) API instead of Playwright, improving reliability and performance. - We improved the [live view](/browsers/live-view) experience to show a clear "disconnected" screen when a browser is destroyed, instead of leaving users on a blank page.

<Update label="January 16" tags={["Product", "Docs"]}>

Product updates

  • Added a Past Sessions view to the dashboard showing your 100 most recent browser sessions; provided context about the session history limit.
  • Improved browser session stability to reduce premature disconnections and WebSocket failures.

Documentation updates

  • Updated Playwright examples on the File I/O page with a more streamlined approach for uploads and downloads.

<Update label="January 9" tags={["Product", "Docs"]}>

Product updates

  • Introduced a new Hobbyist plan, which provides affordable access to features like session replays, uploading and downloading files, and using custom Chrome extensions.
  • Updated the Gemini Computer Use template to use Stagehand V3, resolving compatibility issues.
  • Added support for Actor and Thinker OpenAGI models, with guidance on when to use each approach.
  • Added new Claude Agent SDK templates for TypeScript and Python that demonstrate browser automation using Kernel's Playwright execution API.

Documentation updates

  • Added documentation for the OpenAGI Lux integration using the kernel create method, including examples for both AsyncDefaultAgent and TaskerAgent.
  • Improved file I/O documentation with code examples for uploading and downloading files from remote browsers.

<Update label="December 19" tags={["Product"]}>

Product updates

  • Added a new kernel mcp install command to the Kernel CLI that automatically configures the Kernel MCP server for AI development tools like Cursor, Claude, and Visual Studio Code.
  • Added a new Python template to the Kernel CLI for OpenAGI's Lux computer use model, supporting both AsyncDefaultAgent and TaskerAgent with Kernel browser session management.
  • Improved CLI performance when looking up browsers by using a direct API call instead of fetching the full list of browsers.
  • Fixed bug where headless browsers started with no pages. Headless browsers now initialize with an initial page.
  • Fixed bug where replay start and end times were not being returned by the API.

<Update label="December 12" tags={["Product", "Docs"]}>

Product updates

  • Refreshed our Vercel template with an enhanced UI and AI agent example
  • Deprecated the "persistent" feature from browsers in favor of using extended timeout_seconds with Profiles; added name field to browsers for easier identification
  • Unified browser creation tools by porting Create Kernel App functionality into the Kernel CLI
  • Improved replay videos to include full video length metadata, enabling timestamp navigation and deep linking

Documentation updates

<Update label="December 5" tags={["Product", "Docs"]}>

Product updates

  • Added option to include ephemeral and deleted browsers when listing sessions via API
  • Updated the maximum timeout available from 24 hours to 72 hours of no CDP activity
  • Updated the maximum async invocation duration from 15 minutes to 1 hour
  • Improved error messaging when invalid invocation IDs are provided

Documentation updates

  • Stagehand v3: Added code example for using Stagehand v3 with Kernel browsers
  • Vercel AI SDK integration: Expanded Vercel integration documentation for the @onkernel/ai-sdk package
  • Live view: Added code example showing how to embed browser live views in an iFrame
  • CAPTCHA handling in stealth mode: Added documentation on CAPTCHA handling behavior with Cloudflare challenges