Navide is a desktop application that runs coding agents on your computer. This page explains, without the architecture, where its boundaries are. The full account of what individual features send is in Privacy and Data Flows; the security model and its limits are in the Security policy.
What runs on your machine
Everything that makes Navide work runs locally: the application window, the background service that manages terminals, files and Git, the terminals the agents type into, and the state of your workspaces and pipelines. That background service listens only to the application on the same computer. It is not a server other machines can reach, and it is not designed to be exposed as one.
Navide does not require an account to run, and it does not operate a telemetry service for your projects.
What can leave your machine, and when
Data leaves your computer only when a feature you enabled or used sends it. The main cases:
- The coding agents themselves. Each agent CLI (Claude Code, Codex, and the others) talks to its own vendor exactly as it would if you had started it in a terminal. Navide does not proxy that traffic and does not add to it.
- Cloud AI features you configure, such as inline editing or code review, send the selected code and prompt to the provider you chose.
- Git, GitHub and GitLab operations run through the local
git,ghandglabtools against the remotes you configured. - MCP servers and documentation lookups send whatever those servers' tools are designed to send.
- Update checks contact GitHub Releases with the application version.
- Cross-device messaging, if you sign in and pair your devices, sends sealed messages through Navide's relay. The relay cannot read the message body. It can see which device is talking to which, the workspace label and pane name on each end, and timestamps. Your workspace's full path is never sent; a salted digest that cannot be turned back into the path is sent in its place.
If you do not enable a feature, its traffic does not happen. Turning off or removing the feature stops it.
Where your data is kept
Navide keeps data in three distinct places, with different owners:
- Inside your workspace, in a private folder (
.agent-team/). This holds the project intelligence Navide builds up for you: pipeline progress, run history, task context, agent output, token summaries. It belongs to you as the user of that computer, it is excluded from Git, and it is not a mechanism for sharing state with teammates. Anything that should become team knowledge leaves it only through something you deliberately author — a commit, a document, a pull request. - In the application's own data folder. Preferences and registries: roles, pipelines, recent workspaces, analyzer settings, usage attribution, plugin storage, and any cloud AI keys you entered (stored with restrictive file permissions, which are a courtesy against accidental access, not a vault).
- In the tools you already use. Each coding CLI, Git client, model runtime and MCP server keeps its own configuration, sessions, credentials, logs and caches. Navide reads selected logs from them to attribute sessions and report usage; it does not move or copy their credentials elsewhere.
Your source code, tests and documentation stay where they are: in your repository, shared through Git as before.
Who can instruct whom
- You instruct the agents. Navide starts an agent in a pane at your request and relays what you type. Nothing runs an agent on its own initiative.
- Agents act with your permissions. An agent Navide launches has the same rights on your computer as you do. Navide does not add a sandbox around it; whatever confirmation prompts or sandboxes exist are the CLI's own. If you turn on a mode that skips those prompts, the agent will run commands without asking. Use that only in workspaces you trust and can roll back.
- Agents can message each other, when you allow it. Panes in the same workspace, in another workspace window, or on another paired device can send each other messages. A message is delivered into the receiving agent's input as text; it carries no special authority. What the receiving agent does with it is governed by that agent's own permissions — which are yours.
- Other devices need your say-so. Cross-device messaging works only between devices signed in to the same account and paired by you. A device you have not paired cannot address your panes.
- Plugins are held at arm's length. Extensions such as the Git view run in an isolated environment and reach Git, files and the network only through narrow, host-controlled bridges. They never receive your Git credentials; the host injects a credential for one specific operation and destination at the moment it is needed.
What Navide does not promise
- It does not provide a complete sandbox for agents. An agent can do anything you can.
- It does not guarantee that secrets are scrubbed from text handed between agents, run history, logs or exported diagnostics. Do not put credentials in prompts, plans or files meant for handoff.
- It does not retain your data on a server. The relay stores sealed messages until delivery and the metadata around them; everything else lives on your devices.
These limits are stated deliberately. The Product Roadmap treats policy, isolation and secret handling as long-term work, and this page will change when that work ships — not before.