Strawberry vs OpenClaw

You should not have to become the person who maintains your own assistant.

Picture the job you keep moving to Friday.

Say it was reconciling last month's invoices against what actually landed in the bank, for your finance lead. You need every invoice you raised. You need every payment that cleared. You need the two that do not match.

The information is not missing. It sits in the accounting system, in the bank export, and in a payment thread from March. It is slow because a person has to go and fetch it, one invoice at a time.

OpenClaw starts one step earlier than the job. Its getting started guide wants a supported Node.js runtime and an API key from a model provider. Then you run onboarding, and check that the Gateway is listening on port 18789. The same guide says provider sign-in, channel pairing, daemon install, network downloads, skills or optional plugins can make full onboarding take longer.

Strawberry starts at the invoices. You download the browser, sign in, and ask. Your AI companion opens the accounting system, pulls the bank export, and works through the invoices at the same time. Your finance lead gets the matched rows, the two that do not match, and the evidence beside each one.

The bank statement is behind a login you already have open.

The bank export is not an API call.

It is a page you reach by being signed in, and Strawberry is where that session lives. Your AI companion opens the accounting system, the bank portal and the supplier portal that has no API at all.

OpenClaw's browser documentation describes a different arrangement. Its default is a dedicated Chrome, Brave, Edge or Chromium profile that the agent controls, and it says that profile never touches your personal browser profile. The same page calls it a separate, agent-only browser and says it is not your daily driver.

It documents two ways to reach your real signed-in Chrome instead. One attaches to your session, and Chrome shows a blocking "Allow remote debugging?" prompt the first time, so someone must be at the computer. The other drives your tabs through the OpenClaw browser extension.

With Strawberry there is nothing to attach to, because it is the browser you are already working in. Your connected accounts cover 2,000+ integrations, and the sign-in covers the supplier portals that never built one. The reconciliation arrives as rows in your sheet with the invoice linked beside each value. The question to the supplier sits in your drafts. The ledger correction is prepared against the real record.

Deciding what an assistant may send is not the same as deciding what it may run.

An assistant that can read your bank export and an assistant that can email your supplier are two different amounts of risk. Strawberry treats them as two different decisions.

Every connected action is set to ask, always allow, or off. Let it draft supplier emails all day. Require your approval before any of them send. Switch payments off completely and it stays off.

When something does need approval you get a real editor, not a yes or no box. Before that supplier email moves you see the address it resolved, the full text and the attachments. You can change any of it in place.

OpenClaw documents its controls at a different layer. Exec approvals are managed with its approvals and exec-policy commands, and the presets are named yolo, cautious and deny-all. Pending requests resolve as allow-once, allow-always or deny. Session permission modes set a filesystem boundary and an escalation reviewer, and the full mode is listed as unrestricted filesystem access with no reviewer.

Sandboxing is off by default, and the sandboxing page says it is not a perfect security boundary. The security page describes the threat model as running an AI gateway with shell access. It tells you to run its security audit after any config change, and before exposing network surfaces.

A schedule is only as reliable as the machine somebody keeps running.

The reconciliation is not a one-off. It comes back on the first working day of every month, forever.

OpenClaw has a scheduler. Its automations page describes persisted jobs, run history you can list per job, and delivery to a chat channel or a webhook. The same page says automations run inside the Gateway process, and that the Gateway must be running for schedules to fire.

A Strawberry routine runs daily, on an interval, on a cron expression, or when an app event fires such as a new email arriving. The reconciliation runs at seven on the first working day. The drafts are waiting when your finance lead opens the laptop.

Every run records what it did and the credits it used, so you can audit the work instead of trusting it. Nothing fails quietly either, because a run that nothing picked up is marked expired.

Your finance lead should be able to receive a working AI companion without installing anything.

An AI companion is not an install.

You brief it once on your suppliers, your ledger and the way you write, and it keeps that. Creating a second one costs you a description, not a deployment. The one that chases suppliers and the one that reads contracts are two colleagues, not two hosts.

Then give one away. An AI companion carries its memory, its skills and its instructions, so your finance lead receives something that already knows the invoice book. They sign in as themselves to reach the ledger. Your team workspace holds the shared files and the process, and shared credits cover whoever needs them.

OpenClaw documents a different model. Its multi-user page says everyone who can operate an agent can make it do anything that agent can do. It adds that session ownership and presence are usability features, not security boundaries. People who must not reach each other's sessions, tools, credentials or files are told to use separate agents or separate host trust boundaries.

Frequently asked questions