Anchor Browser vs Kernel (2026)

Managed authentication and authenticated applications give agents logged in sessions on outside sites, with MFA, email one time codes and a 1Password integration for credentials. Named integrations connect Anchor to automation tools (n8n, Make, Zapier), agent and browser frameworks (CrewAI, LangChain, Stagehand, Browser Use) and sandboxes (Daytona, CreateOS, Tenki).

Sourcedocs.anchorbrowser.io/essentials/authenticated-applicationsread 2026-09-23

A named integration catalog puts a Kernel browser inside Browser Use, Stagehand, the Claude Agent SDK and Claude Code, the Vercel AI SDK, Replit, Hermes Agent and computer use models from Anthropic, OpenAI and Gemini. Agents can also act with authentication. Managed Auth logs agents in to outside sites with profiles the user approves, vaults fill stored credentials without exposing them, and wallet integrations with Stripe Link and Agentcard let agents pay.

Sourcekernel.sh/docs/integrations/overviewread 2026-09-27

Automation Tasks are versioned workflows of ordered segments that mix deterministic Playwright code with agentic steps, and they run in Anchor's own runtime. Anchor launches the browser, authenticates, and walks each segment in order, with a workflow graph showing step status. Runs can be synchronous or asynchronous, whole runs retry on a new browser, and a single step can fall back to an AI agent mid-run.

Sourcedocs.anchorbrowser.io/tasks/overviewread 2026-09-23

Customer code runs on Kernel as apps with named actions, invoked synchronously or asynchronously with status streaming, and browser pools keep browsers ready for repeated work. The workflow logic itself lives in the customer's agent, and there is no workflow model for multi step or multi agent work.

Sourcekernel.sh/docs/apps/invokeread 2026-09-27

Anchor describes Automation Tasks as versioned, repeatable workflows with managed identities, scheduling and session control, but no schedule setting is offered. Runs start with an API call (POST /v2/tasks/{taskId}/run), task settings cover identity, human-in-the-loop, retries and browser configuration with no schedule among them, and webhooks send outbound event notifications instead of starting runs.

Sourcedocs.anchorbrowser.io/introductionread 2026-09-23

Work starts when a caller creates a browser or invokes a deployed app through the API, SDKs or CLI, synchronously or asynchronously. No schedule, event or inbound webhook of Kernel's own starts agent work.

Sourcekernel.sh/docs/apps/invokeread 2026-09-27

Anchor keeps no knowledge base or retrieval layer over the customer's own content, and has no index, vector store or grounding layer of its own. Its agents and tools read and extract content from live web pages through research tasks, screenshots and file downloads for the customer's own use, which is extraction from the public web, not search over the customer's documents.

Sourcedocs.anchorbrowser.io/examples/research-taskread 2026-09-23

There is no knowledge grounding or retrieval layer. Kernel supplies browsers and their runtime, and the pages the agent reads are web work done at run time.

Sourcekernel.sh/docs/llms.txtread 2026-09-27

The agent has no memory of its own with a set scope and lifetime. Between runs, Anchor keeps the browser's state. Managed authentication keeps agents signed in to outside applications through encrypted, session-based authentication, and browser profiles carry a stable identity, including a dedicated sticky IP, across sessions. The browser loads each of these at session start, so they work like a saved browser login, not as context the agent reads to decide.

Sourcedocs.anchorbrowser.io/essentials/managed-authenticationread 2026-09-23

Browser profiles save cookies and logins for reuse, Managed Auth keeps sessions signed in, and standby mode keeps a browser ready to resume. Each is browser state loaded or kept for the next session, not a memory the agent reads to decide. There is no memory with a stated scope and lifetime.

Sourcekernel.sh/docs/browsers/profilesread 2026-09-27

With human in the loop enabled, the agent pauses and posts an intervention request to a session-specific queue. The customer lists pending requests and sends a response through API endpoints, or subscribes to the intervention.requested webhook, whose payload carries the request ID, the agent's message and a live-view URL for one-click takeover. Automation Task settings switch human-in-the-loop on per task, so a person can review and direct the agent mid-run.

Sourcedocs.anchorbrowser.io/agentic-browser-control/human-in-the-loopread 2026-09-23

Purchases wait for approval. Wallet integrations supply hosted purchase approval, and for Agentcard, Kernel holds a recognized checkout request for approval before it executes. Managed Auth sends logins through a hosted page where the user approves and enters credentials or MFA, and a live view with read and write access lets a person take over a running browser.

Sourcekernel.sh/docs/browsers/paymentsread 2026-09-27

SOC 2 Type 2, ISO 27001 and GDPR come from the Growth plan, with a BAA, HIPAA and a DPA on Enterprise, and a trust portal holds the compliance report. SSO and RBAC are on Enterprise, and user roles are part of the account configuration Anchor keeps. Each browser runs in a dedicated virtual machine that is destroyed at session end, with strict tenant isolation, encrypted session-based authentication, and a Zero Data Retention mode that turns off all video and log recording for an account.

Sourcedocs.anchorbrowser.io/securityread 2026-09-23

Projects isolate browsers, profiles, credentials and deployments, API keys are scoped to the org or project with expiry and rotation, and Enterprise adds single sign on. Kernel holds SOC 2 Type II, HIPAA, PCI DSS and ISO 27001 attestations, with reports on request through trust.kernel.sh, and GDPR is in progress.

Sourcekernel.sh/securityread 2026-09-27

An Automation Task run shows a steps panel that streams per-segment logs with timestamps, an Agent Logs section wherever an AI agent completed a step, and a workflow graph marking each segment's status. The Runs tab keeps every past execution, and View run replays it against the session recording. Sessions are recorded to MP4 by default and served through the API and web UI, console logs can be turned on, and a session history endpoint lists past sessions. Retention is configurable down to a Zero Data Retention mode.

Sourcedocs.anchorbrowser.io/essentials/recordingread 2026-09-23

A live view, session replays with retention by plan, browser telemetry that can be read or streamed and exported to OTLP destinations, and app logs show what ran. Organization audit logs record every authenticated API request and are kept for one year, with search, download and continuous S3 export on higher plans.

Sourcekernel.sh/docs/info/audit-logsread 2026-09-27

Region choice depends on the plan. Free and Starter run in the US, Growth adds the EU, Asia and Australia, and Enterprise can use any region. By default, infrastructure is hosted in the United States, with customer data processed and stored in AWS us-east-2 (Ohio), and EU-based infrastructure is available for organizations that need it.

Sourcedocs.anchorbrowser.io/pricingread 2026-09-23

A region is chosen per browser or browser pool from us-east, eu-west and ap-southeast on Start-Up and Enterprise plans, with us-east the default. Kernel frames the choice as one of latency, and regional support for apps and invocations is not yet available. Enterprise adds zero data retention, and private networking routes browser traffic to private services, which is connectivity, not deployment.

Sourcekernel.sh/docs/browsers/regionsread 2026-09-27

Starting points come in two forms. Task examples cover form filling, buyer intent, configuration collection and research, and a library of guides shows how to automate named business applications such as NetSuite, SAP S/4HANA, Salesforce and ServiceNow, and US federal forms such as IRS Form W-7, Form I-9 and VA Form 21-526EZ, with Playwright. Each guide is a set of setup steps and a script that a developer adapts, examples to build from, not a catalog of agents ready to adopt.

Sourcedocs.anchorbrowser.io/examples/task-examplesread 2026-09-23

Kernel ships agent skills, including Bot Detection, Browser Profiles, Kernel Auth and Create Site Skills, with a fuller list on skills.sh. They teach the customer's own agent to use Kernel and to turn a verified workflow into a site skill. These are starting points a developer builds from, not a catalog of agents a customer adopts.

Sourcekernel.sh/docs/skills/overviewread 2026-09-27

Customers choose the model behind Anchor's own agent. Perform-web-task takes an agent parameter, browser-use by default, openai-cua or gemini-computer-use, and AI steps are priced by the tier of the model used. Bring Your Own Keys lets a project store its own provider keys for OpenAI, Anthropic, Google, Azure or any OpenAI-compatible endpoint and opt in per request, and Anchor's default model applies otherwise.

Sourcedocs.anchorbrowser.io/advanced/bring-your-own-keyread 2026-09-23

The model lives in the customer's agent, not in Kernel. The computer use guides for Anthropic, OpenAI, Gemini, OpenAGI, Tzafon and Yutori show those agents calling a Kernel browser.

Sourcekernel.sh/docs/integrations/computer-use/overviewread 2026-09-27

Anchor can be called from outside through a REST API at api.anchorbrowser.io, authenticated by API key, that covers browser sessions, session history, AI tasks, automation tasks, agent access and webhooks. SDKs exist for JavaScript (npm anchorbrowser) and Python, alongside a hosted MCP server for any user with an API key and an open source MCP server.

Sourcedocs.anchorbrowser.io/quickstart/use-via-apiread 2026-09-23

Outside code calls Kernel through an API described by an OpenAPI 3.1 specification at kernel.sh/openapi.json, with an API reference at api.onkernel.com. There are official TypeScript, Python and Go SDKs, a CLI, and a hosted MCP server at mcp.onkernel.com over streamable HTTP with OAuth metadata.

Sourcekernel.sh/docs/reference/mcp-serverread 2026-09-27

Automation Tasks are versioned, so a draft is test-run with real inputs before it replaces the live version, and every run carries a status and per-step results. Workflow self-healing analyzes a run whose deterministic step failed and creates a new draft version for the customer to review, turning what the agent completed at run time into a stronger workflow. There are no scored test cases.

Sourcedocs.anchorbrowser.io/advanced/self-healingread 2026-09-23

Replays, telemetry, a browser REPL and debugging skills let developers inspect what an agent saw and did after a run. That is debugging, and there is no evaluation harness, scored test cases or quality gate.

Sourcekernel.sh/docs/browsers/replaysread 2026-09-27

Running real browsers for agents is the core product. Headful Chromium sessions run in dedicated virtual machines that an agent drives through CDP with Playwright or Puppeteer, through Anchor's own browser agent, or through Anthropic, OpenAI, Gemini and Yutori computer-use agents. Sessions come with OS-level mouse and keyboard control, stealth, proxies, CAPTCHA solving and authenticated sessions.

Sourcedocs.anchorbrowser.io/advanced/os-level-controlread 2026-09-23

Kernel runs isolated cloud Chromium browsers that agents drive through a computer controls API (click, drag, type, scroll, screenshots), Playwright execution, CDP or WebDriver BiDi, in headless, headful or GPU modes, with stealth, proxies and CAPTCHA handling.

Sourcekernel.sh/docs/browsers/computer-controlsread 2026-09-27