Watch it do the work.
One machine, in chapters: a chat, its approval, a goal loop, a knowledge search and what it made.
Recorded at the release this site describes.

The tour could not play here. Its chapters, as text:
- Home, and what needs you. Home counts what needs its owner, an approval and a workflow's choice, with an Answer beside them.
- A chat, in your own words. A question in the owner's own words, answered on a local model: it reads the code, runs the tests and explains the failure.
- The approval it stops on. Before it edits a file, the run stops and asks: the tool, what it would change, its risk, and how far an answer is remembered. The answer is the owner's.
- A goal loop and its findings. A goal loop that ran six cycles on its own, and the finding it recorded in each.
- The Library, searched locally. Notes and links kept on this machine narrow as a word is typed.

Ask, then let it do the work.
A chat that can only answer leaves the doing to you. Here a conversation runs tools, writes files and starts longer work, and it stops to ask before a call you have not allowed.
What you control
- Each chat's permission mode: Normal, Trust reads or Trust. YOLO approves everywhere until it runs out.
- Its memory: Persistent, Temporary or Incognito, chosen when the chat starts.
- Its model, chat by chat.
- Where it goes next: branch from a message, rewind a turn, or regenerate the latest answer and switch between its versions.
- The files it makes, kept in a library with every version, to revert or compare.

Hand off a goal. Watch it work.
Long work does not fit in one reply. A goal loop plans, works cycle after cycle and reports against the goal you set, in a cockpit that shows its plan, its current cycle, its tool calls and what it produced.
What you control
- Pause and resume a loop, nudge it while it runs, or stop it for good. A stopped loop keeps its findings, its report and any work not yet merged.
- Whether it asks: an Attended loop stops when it needs an answer, and the question waits in your inbox.
- A kill switch that holds every kind of unattended work at once, subagent spawns included, until you resume it. Your chat turns keep running.
- Daily token and dollar ceilings on what your automations spend.

Stop explaining yourself twice.
Every new chat with a model starts from nothing. Memory keeps what it learns about you: facts, past episodes, the lessons your corrections teach it, the facts it promotes from what keeps coming up, and which tools have served it well.
What you control
- What it holds: see all of it, edit its facts, and delete any fact, episode, lesson or entity.
- Which chats it learns from: a Temporary chat neither reads nor writes it, and an Incognito chat reads it and writes nothing back.

Answers from your own sources.
A model knows the public web, not your notes. Knowledge holds what you add on purpose, kept apart from memory, in a database on your machine, and chat searches it when a question needs it.
What you control
- What goes in: notes, links it fetches when you save them, and files from documents and spreadsheets to images, audio and video.
- What it knows about each item: with a model set, the people, places and things in it.
- Where it lives: one database and a folder of originals in your PersonalClaw home.

Work that starts without you asking.
Recurring work should not wait for you to remember it. Schedules, events and finished runs start work on their own, and what needs you lands in one inbox instead of a toast that scrolls past.
What you control
- When it runs: on an interval, a cron schedule or once, on a lifecycle or data event, or when another run finishes. A schedule can run an agent in the background.
- Webhooks from programs on this machine only. A sender elsewhere comes in through a tunnel or relay you run here.What webhooks can and cannot do
- On or off: every schedule and trigger has its own switch.
- What interrupts you: Notify, Badge, Digest or Never, for each kind of notification.
- One inbox for a loop's question, an approval, a proposed skill or a message. Each row links back to the loop, chat or run that raised it.

Bring the agent that fits the job.
One vendor's agent should not be the only one you can run. PersonalClaw runs its own agent, or a coding agent you install as an app over the Agent Client Protocol, and the models behind them are apps too.
What you control
- Which agent a chat runs on: the native agent, or a coding agent you installed.
- One approval gate for every call an installed agent asks about, and a record of any call it ran without asking.
- Your models: every cloud provider is an app you add or remove. Core ships only a local one and a small offline fallback.
Run it on your machine.
The gateway, the dashboard and everything they keep stay on your machine. No analytics, no crash reporting and no usage reporting; by default it does ask GitHub for a newer release when it starts and on an interval, and one setting turns that off.