Meet BotHearth

An agent with its own computer.
The computer is yours.

Give BotHearth a task. It opens its own browser, works through the steps, and saves the result on your machine — while you watch. When a site needs a real person, you take the keyboard.

  • No vendor quota, no vendor VM. BotHearth runs on your Mac or Linux machine. It signs in to the Codex or Claude Code account on your plan. No BotHearth account to create, and no BotHearth quota to hit. Your own plan is the only ceiling. Codex is verified live. Claude Code is listed, not yet verified live.
  • It stops for you, not at you. It pauses when a site needs a password, one-time code, passkey, CAPTCHA, or card entry. BotHearth then hands you control. The model is frozen while you type and cannot see the screen. Turn on Settings → Ask before sensitive actions for extra checks before sending, uploading, deleting, paying, or new sites. Those checks stay off until you enable them.
  • Long work survives. Our recorded inbox test ran 51 minutes across two runs. It paused partway and picked the work back up. It created 18 filters, archived 21 conversations, and saved one report. Observed result, not a guarantee. See the recorded run and its timings.

Build from source on Mac or Linux with Docker. About 15–40 minutes. Source-available under PolyForm Noncommercial — not open source. Alpha, single operator. One business bot free, permanently. Model access and hosting are separate. The agent browses in its own containerised Chromium; it never touches your browser or your logins.

Ask · Browser · Take control · Saved file
Edited for length. Footage is from an earlier host-native run; current native tasks run inside the computer. Public-page research and a Search Console handoff; no completed Google sign-in. Instrumental audio; no narration. Read the workflow.

You control

The computer stays with you.
The model still sees the task.

Remote models still receive the task context you send.

What you control in this alpha
You controlIn this alpha
MachineDaemon on your Mac, Linux host, or a VM you operate. Browser, shell, and proxy containers. No BotHearth-hosted agent account. Laptop sleep stops local work.
Model accountNative: Codex (verified live) or Claude Code (listed; not yet verified live). API adapters and local endpoints are configured in YAML. BotHearth does not sell tokens.
Data locationWorkspace, task records, audit chain, and browser profile on this computer’s volume. A fresh computer starts clean.
Review promptsOff by default. Optional Settings toggle Ask before sensitive actions reviews detected sends, uploads, deletes, checkout steps, and new-site form submits taken through BotHearth’s browser tools. Expand Sensitive actions to choose which gates fire and to set API-adapter max tool calls and spend cap (0 = no limit); native Codex and Claude Code tasks have no BotHearth cap. Native CLI shell and network tools bypass those checks. Detection is incomplete.
Credentials / takeoverPassword, OTP, passkey, CAPTCHA and payment-card fields pause for Take control. Checkout steps without a card field are optional review prompts. You type; the model is frozen and capture is blocked. Take control requests full screen; Esc leaves full screen without returning control. Passkeys generally fail inside the container browser.
StoppingStop ends the task. bothearth stop stops the daemon. There is no BotHearth quota. A task runs until it’s done, until you stop it, or until your own model plan hits its limit. Then it pauses and can resume. Closing the window is not Stop.
NetworkDefault: public web via the egress proxy. Strict mode restricts destinations. Policy is best-effort, not a firewall.

Data privacy · Visible work · Human control

Choose where it runs.
Stay involved in what it does.

Your host stores the primary workspace, task records and browser profiles. Remote models still receive the context you submit. Know the boundary before handing over a task. The agent browses in its own Chromium inside a container. It never touches your browser or your saved logins.

YOUR INFRASTRUCTURE

Keep the workspace with you.

Run on your Mac, Linux computer or a prepared Linux VM. Choose the host, storage, network access and backups. BotHearth does not require its own hosted agent account.

YOUR MODEL ACCESS

Connect Codex or Claude Code.

Codex is verified live. Claude Code is listed, not yet verified live. API adapters and local endpoints are a separate YAML path. Provider terms, processing and charges still apply; self-hosting is not a promise of zero data egress.

YOUR CONTROLS

Watch, take over, optionally review.

See the browser. Take control when a site needs you. Enable optional review prompts in Settings for detected BotHearth-tool actions. Pause or stop work.

A recorded test

A real inbox.
A real organisation run.

Private details blurred. Not an independent benchmark.

BOTHEARTH AT WORK Actual product capture
BotHearth organising a real Gmail inbox during testing, with account identity, senders and message content blurred for privacy.
Real inbox. Real email organisation test. 51 min 6 s across two runs. Private details blurred. See the completed result & test details.

What will you ask it to do?

Live websites.
A computer for the work.

A topic to research, current options to compare, or a website to improve. Give it the pages and say what you want back.

A task, from start to finish

You give it the task.
It works through the steps.

BotHearth puts the request, the agent’s actions, and its saved files in one place. Here is a small task you can try yourself.

Illustrated workflow · audit your public website

01 · YOU ASK

“What does my website leave unclear?”

Browse my public pages, identify unanswered customer questions, and suggest changes I can review.

Try the website audit
02 · THE AGENT WORKS

You follow along.

  1. Opens the live pages
  2. Checks the visitor’s questions
  3. Saves findings with source links
03 · YOU REVIEW

website-audit.md

Five changes to review

Affected page · observed issue
Suggested change · how to check it

Follow the links and verify each finding

Review the saved audit against your live pages and product before publishing any change.

WATCH

See its computer.

Follow the live browser beside a feed of what the agent opens, reads, and saves. Check where it is headed as the task unfolds.

STEP IN

Take the controls.

Enable optional review prompts in Settings. When a site needs your sign-in, take control yourself (full screen; Esc leaves it), then give it back to continue.

CHECK

Open what it made.

When a task saves a file, open it from the result view. Read the summary, inspect the source, or check the numbers before using them.

Limited-time enterprise offer

Your first business licence.
Claim it free. Keep it forever.

One perpetual business bot per organisation. Sell the work it creates. Additional bots: US$99 US$49 each, once. Limited-time introductory offer — over 50% below the planned standard price. Payment collection is not open yet. Commercial outputs are allowed; hosting a copy of the service requires a separate agreement. The published first-installation grant needs no registration and survives withdrawal of the offer for future grants. Certificate issuance is not open yet.

Explore the enterprise offer →

Your next task starts here

Give your agent
something to do.

Set up BotHearth, connect your model account, and try a task of your own.

Build from source on Mac or Linux with Docker. About 15–40 minutes. Free under the business offer. Model access is separate.