Sigma Browser

The agent works with websites the way a person would, opening pages, clicking buttons, filling fields, submitting forms and gathering data, so it reaches sites and spreadsheets through the browser itself. No connectors, custom tools or framework for authenticated actions in other systems are offered alongside it.

SourceSigma Browser, sigmabrowser.com/llms.txt and /sigma-ai-agentread 2026-10-06

The agent takes a stated goal and works through multi step website tasks, carrying page context from one step to the next inside an agent workspace that brings summaries, comparisons, research, planning and browsing together. It handles forms, page steps and site flows that would otherwise take manual clicks, but each task is one agent following one goal. Sigma names no way to compose several agents or chain tasks into a workflow.

SourceSigma Browser, sigmabrowser.com/sigma-ai-agentread 2026-10-06

Deep Research lets users queue sources, cluster information and generate structured insights for each request, and Chat with Page answers questions about the page, PDF or document in view. The agent can also summarize pages and review uploaded files, but Sigma names no maintained index over a user's or team's own documents. Knowledge comes from whatever is open or fetched for each task.

SourceSigma Browser, sigmabrowser.com/llms.txt and /sigma-ai-agentread 2026-10-06

Before the agent submits a form or clicks through a site, Sigma names no approval step, and it offers no limit on the actions the agent may take or way to pause a task and take over. The browser includes prompt injection defense, meant to keep the pages the agent visits from steering it.

SourceSigma Browser, sigmabrowser.com and /sigma-ai-agentread 2026-10-06

Sigma states that all user data and AI chat conversations are end to end encrypted, that agent tasks are processed with end to end encryption, and that credentials and task data are not stored after execution. The browser code is public so its security can be reviewed, phishing sites are detected in real time before the user lands on them, and the company says it complies with GDPR and CCPA. There are no user roles, single sign on or admin controls for a team, and no SOC 2, ISO 27001 or other independent audit is published.

SourceSigma Browser, sigmabrowser.com and /llms.txtread 2026-10-06

The agent offers no run history, action log or trail of the steps it took on each site for the user to review afterward, and Sigma states that credentials and task data are not stored once a task has run.

SourceSigma Browser, sigmabrowser.com/llms.txt and /sigma-ai-agentread 2026-10-06

Credentials and task data are not stored once a task finishes, and Sigma names no memory the agent carries from one task to the next. Private profiles keep browsing spaces separate from one another, and the agent does not draw on them as memory.

SourceSigma Browser, sigmabrowser.com/llms.txt and sigmabrowser.comread 2026-10-06

In Private Mode the agent and chat run on local models on the user's own device, with prompts, pages and files processed there and nothing sent to external servers, and they work offline with no connection needed. Modern laptops, including Apple Silicon Macs, run the compact local models comfortably. The browser itself is a desktop and mobile client for macOS, Windows and iOS, with Android and Linux versions on the way, and its cloud features offer no choice of region or data residency.

SourceSigma Browser, sigmabrowser.com/local-ai and sigmabrowser.comread 2026-10-06

OpenClaw and Hermes are general purpose open source agents that take whatever goal the user gives them, and no catalog of prebuilt agents or templates for set jobs is offered. Agent use is limited on the free plan, Plus adds five more agent uses, and Pro makes the agent unlimited, with Deep Research, image generation and web search allowances rising across the same plans.

SourceSigma Browser, sigmabrowser.com/sigma-ai-agent and /plansread 2026-10-06

Every agent task starts when the user switches to Agent mode in the browser and states a goal. Sigma names no schedule, event or background trigger that starts the agent on its own.

SourceSigma Browser, sigmabrowser.com/sigma-ai-agentread 2026-10-06

The user chooses which agent and which model run each task, picking OpenClaw or Hermes in Settings and then connecting an API key for a cloud model or selecting a local model in Private Mode. Local models download in the browser and run on the device, and chat can switch between local models, including the built in Eclipse model, and cloud models at any time.

SourceSigma Browser, sigmabrowser.com/sigma-ai-agent and /local-airead 2026-10-06

Sigma offers no API, SDK or MCP server for driving the browser or its agent from another program. The browser source code is public and the Eclipse model has its own repository on GitHub, so developers can read and review both. The API key a user connects lets the agent call that user's chosen model provider.

SourceSigma Browser, sigmabrowser.com and /local-airead 2026-10-06

The Speedometer 3.0 benchmark compares how fast the browser runs against other AI browsers. It measures page speed, not the agent's work. Sigma names no test cases, scoring of agent tasks or check on the agent's results.

SourceSigma Browser, sigmabrowser.comread 2026-10-06

OpenClaw or Hermes runs inside Sigma's own Chromium based browser, where the agent opens pages, reads content, clicks buttons, types into fields and works through multi step website tasks on the live page. The agent is chosen in Settings and ships with the browser, so it needs no setup or coding, and it can summarize pages, compare product pages, fill in fields and turn long content into useful next steps.

SourceSigma Browser, sigmabrowser.com/sigma-ai-agent and sigmabrowser.comread 2026-10-06