Settings
Local data and privacy
Where LoCrew stores your workspace, how API keys are encrypted with the operating system's keychain, what never touches disk, what leaves your computer, and how to back up safely.
LoCrew is local first: the app, its database and your credentials stay on your computer. This page lists exactly what is stored, where, and what leaves the machine.
#The data folder
Everything lives under one folder. Its path is shown in Settings → About → This installation → Data folder.
| Item | Contents |
|---|---|
locrew.db | The SQLite database: agents, channels, direct messages and messages, runs and their recorded activity, flows and their versions, runs and approvals, providers, MCP servers, tool grants and settings. |
attachments/ | Images sent with messages, one folder per conversation. |
skill-plugins/ | Skills used by Claude Code agents. |
If you used LoCrew when it was called AgentWorkspace, it keeps using that older folder and database so nothing is lost.
The default working directory for new agents is your Documents folder. Agents read and write in the folders you give them, not in the data folder.
#Credentials and the keychain
API keys, MCP tokens and environment values, sensitive headers, A2A credentials and Flow credentials are stored encrypted with your operating system's keychain:
| OS | Mechanism |
|---|---|
| Windows | DPAPI |
| macOS | Keychain |
| Linux | libsecret or KWallet |
Keys never reach the user interface, are never logged, and are stripped from error messages. Header names that look like credentials (containing authorization, api-key, token, secret, password, cookie or session) are always stored encrypted.
If encryption is not available, saving fails rather than falling back to plain text:
Secure credential storage is not available on this system, so the key was not saved. On Linux, install and unlock a keyring (GNOME Keyring or KWallet), then try again.
Moving a profile to another computer
An encrypted key can only be read on the computer and account that encrypted it. If you copy the data folder elsewhere, stored keys read as empty. Enter them again.
#What is never written to disk
- Work sessions (temporary write access) are held in memory. Closing LoCrew always returns to asking.
- Agent gateway tokens, the credentials agents use to call LoCrew's own tools, are created for each run, held in memory, and retired when the run ends.
- Sidebar width and whether it is collapsed.
#What leaves your computer
LoCrew itself has no account and no server of its own. What leaves the machine is what your runtimes send:
| Runtime | Where the AI runs |
|---|---|
| Claude Code, Codex | At Anthropic or OpenAI, under your own login, exactly as when you use those CLIs directly. |
| Hosted model providers | On the provider's servers, under your API key. |
| Local model servers (Ollama, LM Studio, vLLM, llama.cpp) | On your machine. |
| External A2A agents | Wherever that agent runs. |
The conversation content an agent sees reaches its runtime under that runtime's terms. HTTP steps in flows also send requests to the addresses you configure.
Important
File access levels are limits, not a sandbox. Claude Code and Codex run as your user account. Use a virtual machine or container if you need real isolation. See Agent permissions.
#Backing up
- Close LoCrew, then copy the whole data folder. Copying a database file while it is open can give you an inconsistent copy, because LoCrew uses SQLite's write-ahead log.
- Database migrations only move forward. Back up before trying an older version of LoCrew.
- After restoring on a different computer or user account, re-enter your API keys.
#Startup housekeeping
Every launch, LoCrew tidies up after a crash or quit:
- Runs left in progress are marked failed, and their activity reactions show as interrupted, never completed.
- Flow runs that were cut off are marked Interrupted. You choose whether to resume them.
- Old flow step inputs and outputs are cleared according to Flow step history.
- Agents whose runtime is found are shown as Ready; the others as Unavailable. This check does not call a model.
- MCP servers set to Connect on startup connect in the background. Local servers whose command you have not approved wait for you to press Connect.