What leaves your machine
Nothing leaves unless you configured the thing it leaves through. Everything below is nameable,
listable and revocable. Settings → What leaves in the app lists every host a team's agents may
reach, from the teams' own access.yaml files.
What the runtime sends
The agents run on your computer through Claude Code, goose or Codex, and each talks to its own model provider under your own account. That traffic is the runtime's, on your terms with that provider. Agent Teams starts every runtime with its vendor's telemetry opt-outs set.
With local models, which is where agents run when Claude Code is not installed, they go through goose and Ollama on this machine, and model traffic does not leave it.
What Agent Teams sends
- No product telemetry. None is collected today. If it ever is, it will be opt-in and it will never carry source, diffs, prompts, file paths, or repository or branch names.
- A push and a pull request when a task finishes, on its own, or once you approve it if the
team's policy gates
push. Through your owngitandgh. - GitHub calls, once you install the integration.
lucitra integration-add github. The daemon makes them with your ownghlogin, only for agents whoseaccess.yamlallows the tool, andapi.github.comandgithub.comjoin the list of hosts.lucitra integration-remove githubtakes them away from every agent. - Brokerage, market-data and research calls, once you install them.
alpaca(paper unless you install it aslive),market-dataandresearch. Their hosts join the list the same way. - Reads of your own secret sources. Keys for those integrations come from where you already keep
them: GCP Secret Manager through your own
gcloudlogin, AWS Secrets Manager through your ownawslogin (an SSO profile works), a 1Password vault through your ownopsign-in, the macOS Keychain (which stays on the machine), or an env file you name (lucitra secret-source-add gcp|aws|1password|keychain|env …). Agent Teams keeps no copy, and listing which keys are set reads no value. - Model calls, once you point an agent at a provider with your own key. For OpenAI, Anthropic or
an OpenAI-compatible endpoint, the daemon forwards the agent's model calls from
127.0.0.1to that provider and adds your key there, so the key stays out of the agent (lucitra model-providers). The prompts and code an agent sends its model go to that provider. - Agents' own notifications, once you install
ntfy. An agent grantedntfy:notifypushes to your verified ntfy channels, never SMS or WhatsApp, at most 20 an hour per team. - Messages to a channel you added.
lucitra add-channel ntfy|sms|whatsapp, used for gates, finished and failed tasks, and budget alerts, only once you have verified the channel.lucitra revoke-channelstops it.
What agents cannot reach
- The daemon's own credentials. Channel credentials (Twilio, ntfy) are stripped from every agent's environment.
- Anything a policy gates.
deploy,publishand the like are refused in the shell, and a gated integration tool waits for you. - Any integration credential. The daemon makes every integration call for them: the GitHub token
stays in
gh, and API keys are read by the daemon for one call and never enter an agent's environment, prompt, log or command line. - Your model keys, for the providers above. An agent holds a token that works only through the daemon's local proxy and stops working when its run ends. A Bedrock API key is the exception and does reach the agent; an AWS SSO profile avoids handing one over.
What is stored, and where
~/.lucitra (or LUCITRA_HOME) | the store: tasks, threads, every step each run took (kept 30 days), token counts |
~/.lucitra/worktrees | one git worktree per task, outside your checkout |
| your repository | the project memory, as markdown under .lucitra/ you can read, diff and review. Agents propose changes to it as pull requests |
The app
lucitra bridge listens on 127.0.0.1 only, behind a token printed for that run. It makes no
network request of its own and loads nothing from the internet.