Hangar desktop app

Hangar is a desktop app for running and steering Claude Code sessions on a person’s computer, inspired by Claude Desktop. Claude Code ships inside the app. Connected to an Agent Hangar platform, it:

• signs in with the person’s company account;
• can use the token models the admins sync;
• shows the member’s workspace (home, their agents, decisions and inbox) with live updates.

The platform side of the contract is in app-hangar.md (sign-in and devices), app-hangar-models.md (synced token models) and app-hangar-workspace.md (workspace and events).

Installing

Download the app from the Hangar releases:

• Windows: installer or portable .exe;
• macOS: .dmg;
• Linux: AppImage or .deb.

The builds are not code-signed yet:

• Windows: SmartScreen warns about the app; choose More info → Run anyway.
• macOS: right-click the app in Applications and choose Open.

The app’s interface is in Portuguese for now.

Connecting the app

1. In the app, open the settings, enter the platform URL (the same address as the console) and choose Connect.
2. The browser opens the platform’s sign-in and consent screen. The app asks for the app scope (its own API) and the mcp scope (the platform tools). After consent the app receives an OAuth token on a loopback address and registers the device.
3. The app sends a heartbeat while it runs. Revoking the device, revoking the app in API keys → Connected apps or deactivating the person disconnects the app right away.

Admins can set the platform for everyone with a Windows policy, which takes precedence over the app settings: HKLM\Software\Policies\Hangar (then HKCU\Software\Policies\Hangar), values PlatformUrl and Mode (off, optional or required). With required, the app creates no sessions until it is connected.

The app scope opens /api/app/* and an explicit list of member routes, always with the person’s own access; it never opens admin, settings, keys or OAuth routes, and /mcp needs the mcp scope. MCP clients that ask for no scope (Claude.ai, ChatGPT) keep getting mcp tokens, as before.

In the console

App Hangar (admins) has two pages:

• Overview: connected people and devices, the platform URL to give people, how to connect, the Windows policy and the app versions in use.
• Devices: every registered device with its person, version and last heartbeat, and a Revoke button (it also ends the device’s sign-in).

Synced token models

People can run a session either on their own Claude subscription or on a token model the company pays for. To offer a model in the app:

1. In LLM Providers, open an LLM connection that speaks the Anthropic protocol (the only format Claude Code uses).
2. Choose Sync token models with Hangar desktop, then hide any models people shouldn’t pick. With the code policy set to approved connections only, the connection must allow code.

The app calls the models through the central’s LLM proxy (/api/app/llm/{connection}/v1/messages) with a key bound to the device. The provider key stays in the platform’s vault, and the proxy enforces:

• hidden models;
• unsynced connections;
• the code policy;
• team and company budgets.

Each call is recorded in usage with its tokens (cache included) and cost. A person in several teams is billed to their first team. Request and response contents are never logged.

Workspace and live events

The app’s home and My agents use the same data as the member portal. GET /api/app/events streams changes to the inbox, decisions, agent status and the model list over Server-Sent Events, and the app refreshes when it reconnects.

Behind a reverse proxy, turn buffering off and allow a read timeout of at least 60 seconds on /api/app/events (the central already sends X-Accel-Buffering: no for nginx).

Upgrading

Version 0.53 adds migrations 0052, 0053 and 0054, which run on startup. Back up the database first, as for any upgrade (see backup.md).