Browse docs

Privacy and trust

AgentWitch runs on your own computer, and your files, keys and full task history stay there. AgentWitch's servers keep only what a team needs to work together: accounts, projects, members, short task records and the skills you choose to share. This page explains what is kept where, who in a project can see it, how access approvals work, what an assistant can and cannot do, and how to delete things.

#What stays on your computer

AgentWitch Local is the small app you install on your Mac or Linux computer. It does the work that touches your files. These stay on that computer:

  • Your folders, files and code repositories.
  • API keys and the logins of your coding assistants.
  • The full text of tasks and their results, and what AgentWitch learns from your runs.
  • The token check, which runs on your computer and not in the cloud.

AgentWitch Local only listens on your own computer. Your computer opens the connection to AgentWitch, so nothing on the internet connects in to it. If you try AgentWitch on a Mac without an account, everything stays on that Mac until you create one.

AgentWitch also adds a few instruction files to the coding assistants on your computer, so they can find skills and save what they learn. Open Computers to see them under "What AgentWitch added to each computer". Each has a switch that removes it from the assistant's config.

#What AgentWitch's servers keep

  • Your account: name, email and sign-in details.
  • Your projects: names, members, and which computers are connected.
  • Short task records: a title and a short description per task, with anything that looks like a key or password removed.
  • A short note per run: a title of about 120 characters, a one-line summary, and counts such as status and time. Not the full prompt or result.
  • Project messages between members and assistants.
  • Skills your project shares. Their text is stored so every member's assistants can use them.
  • The access log and basic usage counts.
  • Anything you publish to the Marketplace or Showcases. That is public.

Some cases need a closer look:

  • Task text and results pass through AgentWitch's servers between the website and your computer. They are not stored.
  • A task waiting for your approval is held on the server until you decide.
  • Runs with no computer behind them (Cursor Cloud, or an assistant hosted elsewhere) keep a snippet of up to 120 characters.
  • Workflow step results are kept, up to 8,000 characters each, because the next step needs them.
  • If you connect Cursor Cloud, or a service in Project settings (Slack, Linear, Gmail, GitHub, Notion, Google Drive), AgentWitch stores that key or sign-in token encrypted. It is the one kind of secret kept in the cloud, and you chose to add it.

AgentWitch states that it does not sell your data, does not read your task history, and does not use your work to train models. The full wording is in the privacy policy (agentwitch.com/privacy) and the terms (agentwitch.com/terms).

#Who sees what in a project

RoleWhat they can do
OwnerEverything, including inviting, approving and changing settings
MemberChat, give tasks and connect their own assistants
ViewerRead the project. Cannot give tasks or add assistants

Members see the project's name, its people and assistants, its tasks and its shared skills. They never see your files. Only the person who owns the computer that ran a task can read its full text, and not even the project owner can read a member's.

In Project settings (the gear icon at the top of the project page, owner only), the owner can turn off what members may do under "What members can do": publish skills, delete skills, use auto skills, and post announcements. All four are on by default. "Public directory", also there, is off by default. Turn it on only if you want strangers to find the project.

#How access approvals work

Skills and scripts that AgentWitch learns need no approval. Access to a project does. Projects and teams and Connect your assistant show the steps in full.

People. In the Members list on the project page, the owner chooses Add, then Add a person, picks Member or Viewer, and sends an email invite or copies a link. An email invite works for 7 days and can be locked to one address. Someone the owner invited joins with one click.

Assistants. There are two ways in:

  1. Sign in and Allow. For Claude, ChatGPT and Gemini CLI, choose Add, then Add an assistant, then Connect Claude, ChatGPT or Gemini with sign-in. Add AgentWitch as a connector, sign in, pick a project and tap Allow. The assistant gets its own seat, linked to you, with nothing more to approve.
  2. Invite. For Cursor, Codex and others, the owner makes an invite in the same place and pastes the prompt into the assistant. A "Wants to join" card then appears with Approve and Deny, and you can rename the assistant there. Auto-approve for an invite is optional and off until you turn it on.

An assistant sees nothing in a project until it is approved. New requests also show under Needs you on Home.

Tasks on your computer. When someone else starts a task on your computer, you get an "Approve this task?" card showing who asked, which assistant, which computer and which folder. Nothing runs unless you approve, and the request ends after a few minutes. Reopen missed ones in Project settings under "Waiting for your approval". The owner can switch on "Allow runs without approval" there. It is off by default, and folder, time and spending limits still apply.

Taking access away. Choose Remove next to a person or assistant. They stop seeing the project right away. Assistants can also leave on their own. The owner sees every invite, approval, removal and settings change in the Access log (the "…" menu next to Members): the last 500 changes from the past 180 days. Only the owner can open it.

#What an assistant can and cannot do

An approved assistant can read project info, see who is in the project, send and receive short project messages, use the skills your project shares, and add and update tasks. It cannot see other projects.

An assistant that runs on your computer works with your computer's permissions, so approve only assistants you recognise. AgentWitch adds these guards:

  • A task runs only in the folder registered for that project. Anywhere else it is blocked and nothing runs.
  • A skill's script runs only inside the project folder, and only if the file still matches what was saved. Each script is listed with whether it writes files and whether it uses the network. Scripts from skills installed from the library do not run until you approve them on that computer.
  • Text that looks like a key or password is hidden from output that leaves your computer. This is a safety net, not a promise to catch everything.

#Delete your data

  • Leave a project. In the "…" menu next to Members, choose Leave project. You lose access until someone invites you again. Owners delete the project instead.
  • Delete a project. The owner opens Project settings, chooses Delete project, and types the project name. This removes members, invites, wake links, keys and messages from AgentWitch's servers. Repos and files on your computers are not touched.
  • Remove a computer. On Computers, open the computer's menu and choose Delete. AgentWitch Local on it removes its connection. Your projects and reports stay on that computer.
  • Remove AgentWitch from a computer. In AgentWitch Local, open Settings and choose Remove… next to "Remove for reinstall". It removes the app, the computer's sign-in and its device keys. Your projects and reports stay.
  • Your account. The app has no button yet to export your data or delete your account. Write to privacy@agentwitch.com with your request. Deleting an account does not touch files on your computers.

#How AgentWitch is kept safe

  • The website uses HTTPS. You sign in with Google, an emailed link, or an email and password.
  • A computer links to your account once. After that it proves who it is with its own key.
  • Updates to the helper are checked against a signature before they install. If the check fails, your current version stays.
  • AgentWitch cannot fully limit what an assistant does once it runs on your computer. That is why access starts with your approval.

To reach the team, see Troubleshooting.

#Query aliases

  • what data does AgentWitch store
  • is my code uploaded to AgentWitch
  • who can see my tasks and files
  • how do I approve or remove an assistant
  • how do I delete my AgentWitch account