Where does BotHearth run?
The daemon runs on your macOS or Linux host. You open its interface in an ordinary browser. Each agent computer has separate browser and shell containers, a workspace, and an outbound proxy. The model connection is a separate host process or configured API adapter.
- YouBrowser interface
Task, approvals, live view, takeover
- Your hostBotHearth daemon
Policy, task records, vault, MCP tools
- Its computerBrowser + shell + proxy
Chromium, workspace files, permitted destinations
The installed model CLI talks to its provider and uses the daemon's task-scoped MCP tools. The daemon dispatches computer actions through container stdio, with no published sandbox ports and no Docker socket inside the managed computer.
The browser owns a persistent profile. The shell cannot mount that profile; it sees the configured workspace. The workspace is a folder on the host, so saved task files persist there. The container boundary does not restrict the separate host-side CLI or trusted connectors you install.
How do I connect Codex or Claude Code?
For the normal task interface, install and authenticate the official CLI, then select it in Settings → AI connection. BotHearth invokes the installed CLI using its native authentication and supplies task-scoped computer tools.
| Path | Use it for | Important boundary |
|---|---|---|
| Task interface | Starting tasks through BotHearth with installed Codex or Claude Code. | Your provider account, terms, usage limits, and CLI host permissions apply. |
| Manual MCP integration | Connecting a supported external harness to the computer tools. | Requires an operator-created task binding as well as an MCP token. |
| Standalone adapter | A configured API endpoint; also required for scheduled tasks in this release. | Model/tool-call support and price estimates depend on the adapter and endpoint. |
Manual configuration writers exist for Codex, Claude, Gemini, Cursor, OpenCode, and Copilot. A writer produces configuration; it does not prove that a particular installed version or account can complete a task. Follow the manual integration guide for the binding and token steps.
Standalone adapters are an advanced path. API adapters currently drop image content, so do not assume visual-task parity with the native CLI path. Consult provider requirements and adapter documentation before choosing an endpoint.
What happens when I start a task?
- You describe the task. The daemon creates task state and an allowance, and prepares the computer.
- The model chooses a tool. BotHearth validates the task, arguments, policy, control state, and remaining allowance.
- The computer acts. A browser or shell tool runs and returns a result. New destinations and detected sensitive effects may request an operator decision first.
- You inspect the work. The interface shows activity, requests, live view, and saved outputs. Successful process exit alone is not task completion: the model must use the completion tool.
Ordinary permitted interactions and writing into the task workspace can proceed without another approval. A page may contain misleading instructions; classification and model judgment can fail. Read exactly what approvals cover.
What happens when I take control?
BotHearth drains in-flight agent actions and blocks ordinary agent tools before acknowledging human control. Live frames go to the authenticated operator. They are not routed to model results or retained in task or audit records during takeover.
You can type a password or code into the site while controlling the browser. The daemon transports that input and the site receives it. Never put credentials in the chat. Later page content may become visible to the model after you return control.
Give control back validates the page before capture resumes. A remaining sensitive-field signal can block handback. Ten minutes without input pauses control instead of automatically handing it to the agent. Passkeys and hardware security keys generally cannot be completed inside the container's Linux Chromium.
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.
Task steps and estimated usage can pause work. The harness dollar meter counts computer-tool calls; API estimates use configured prices. Neither is a provider-enforced spending cap. Container hardening and approval gates reduce exposure but cannot make arbitrary unattended browsing safe.
Next: install BotHearth, review security boundaries, or try a task with a known check.
Implementation references: architecture, MCP integrations, security model. Reviewed 2026-09-08 for v0.0.1 alpha.