Give it a task.
Follow the work.

From a request to a result

1. Say what you want done. Start with a clear task, the pages or information it should use, and the result you want back.
2. Watch the agent work. Its computer can open pages, read information, and save files. The activity feed shows the steps alongside its live browser.
3. Step in when it needs you. Take control for a password, OTP, passkey, CAPTCHA, or a payment-card field. Optional review prompts in Settings cover detected BotHearth-tool actions, including checkout steps without a card field.
4. Open the result. When the task saves a file, you can open it from the result view. Check the contents and sources before relying on it.

Review prompts are off by default. Settings → Sensitive actions can arm per-gate prompts and set API-adapter max tool calls and spend cap (0 = no limit); native Codex and Claude Code tasks have no BotHearth cap. Ordinary browsing and saving a result proceed without a destination prompt. Native CLI commands have their own file and network access inside the computer; MCP review-prompt detection does not cover those commands. See what optional review covers.

Ready to try it? Choose live web research, a comparison, or a website audit.

Take over when a site needs you

Select Take control to use the agent’s full desktop, including its browser, Files and Terminal. Take control requests full screen; Esc leaves full screen without returning control. You can complete a sign-in or another step that needs a person, then select Give control back to continue the task.

Before control is acknowledged, guest model processes freeze and MCP computer actions and capture are blocked; the operator desktop stays usable. The site receives what you type and BotHearth transports that input; never put passwords in the task text. Messages to the frozen model wait until you hand control back. Later page content and shared workspace edits can become model-visible.

BotHearth checks the page before returning control. A remaining sensitive-field signal can prevent handback. Ten minutes without input pauses control rather than automatically returning it to the agent. Passkeys and hardware security keys generally do not work inside its browser. Read the full control boundaries.

Connect your model account

The computer image includes the official Codex and Claude Code CLIs. Codex is verified live; Claude Code is listed and not yet verified live. In Settings → Model connection, choose one and follow its sign-in instructions. Its login stays in the bot’s computer; host credentials are not copied.

Your account’s eligibility, usage limits, and charges apply. BotHearth does not include model access and is independent of OpenAI and Anthropic. API adapters and local endpoints are a separate YAML path. The setup guide walks through connecting an account; provider requirements cover the details.

Already using another agent interface? The MCP integration guide explains how to connect it to BotHearth’s computer tools. Custom API connections and scheduled tasks use the separate standalone adapter path.

Its computer runs on yours

You run BotHearth on your Mac or Linux computer and open it in a browser. Your agent gets its own browser profile and a workspace for its files. Its saved results stay on that host, where you can open them after the task.

The agent browses in its own Chromium inside a container. It never touches your browser or your saved logins. Containers separate the agent’s tools from your everyday browser. New native model sessions run inside the computer with a shared workspace and private model home. Historical host sessions keep their original host login and conversation. Remote models still receive the task context they need; running BotHearth locally does not make it offline.

Read the security and privacy guide for the boundaries, or the architecture documentation for the implementation.

What should I expect from this alpha?

BotHearth is intended for one operator and source installation on macOS or Linux. It has no hosted agent service, published npm install path, prebuilt images, or notarized desktop app. Websites can block automation, and third-party rules remain your responsibility.

There is no BotHearth quota. API-adapter tasks can set max tool calls and a spend cap under Settings → Sensitive actions (0 = no limit). A task runs until done, until you stop it, until a set cap is reached, or until your model subscription’s usage limit is reached; then it pauses and can resume. The receipt estimates MCP computer-tool calls, not every native command or model request; API estimates use configured prices. Neither is a provider bill. Container hardening and optional review prompts reduce exposure but cannot make arbitrary unattended browsing safe.

Next: install BotHearth, review security boundaries, or try a task with a known check.