# Claude Desktop sign-in — shipped fixes and planned work

Two related sign-in problems when Claude Desktop runs inside the LinuxServer Selkies
streamed desktop.

**Status:**
- **Shipped:** Problem B (sign-in persistence) — the `autostart` bootstrap landed in v1.4, and
  the `gnome-keyring` package itself was added in v1.17. That approach was later reverted:
  v1.24-ish removed `gnome-keyring` again because it prompts for a keyring password on first
  boot, which blocked Claude Desktop from ever launching — but the launch flag was left
  forcing the now-daemonless libsecret backend, so `safeStorage` silently went unavailable
  again (recurring "sign in again", and the Claude app's dispatch tab showing the desktop as
  offline until a fresh sign-in was done from a computer). v1.35 switched to
  `--password-store=basic` plus a cont-init script that re-syncs the persistent openbox
  `autostart` from the image on every boot — **but that flag alone does nothing**, and the bug
  survived it untouched. v1.37 added the application-side opt-in the `basic` backend requires;
  see "Why v1.35 did not work" below. That patcher then regressed silently when upstream's
  bundle output changed shape — fixed again in 07308543.1; see "Why v1.37 stopped working"
  below.
- **Planned only:** Problem A (in-desktop browser for OAuth) is intentionally not implemented.
  The image ships no browser; complete the login with the user-side workaround below.

---

## Problem A — Cannot complete login

Symptoms:
- **"Continue with Google"** does nothing.
- **"Continue with email"** sends a link that "must be opened from the same machine".

### Root cause
Claude Desktop signs in through an **external web page** (Google OAuth or the email magic
link) and then hands the token back to the app via the **`claude://` URL scheme**
(auto-registered on first launch). Both paths need:
1. a **web browser inside the streamed desktop** to render the auth page, and
2. a **default-browser association** so the app's `xdg-open <url>` call goes somewhere.

The Selkies base image ships **no browser** and no default handler, so `xdg-open` has
nothing to dispatch to → Google "does nothing", and the email link's final `claude://…`
redirect only works on the machine running the app (the container), which currently can't
render/handle it.

### Fix (image level)
In `claude_desktop/Dockerfile`, install a browser and register defaults so the whole flow
completes inside the session:
- `apt-get install -y --no-install-recommends chromium` (Debian bookworm package name;
  `firefox-esr` is an alternative).
- Set the default browser and confirm the scheme handlers, e.g. as an s6 oneshot or in the
  desktop `autostart` (runs as the session user with `$DISPLAY`/dbus available):
  ```sh
  xdg-settings set default-web-browser chromium.desktop
  xdg-mime default claude-desktop.desktop x-scheme-handler/claude
  ```
- Verifying expectations (for reference, not a test step):
  `x-scheme-handler/https` → `chromium.desktop`, `x-scheme-handler/claude` →
  `claude-desktop.desktop`.

Then login is done entirely in the streamed desktop: **Continue with Google** opens
Chromium in-session and the `claude://` redirect lands back in the app; or paste the email
magic link into the in-session Chromium (not a phone).

### User-side workaround (no rebuild)
- Add-on Configuration → `additional_apps: chromium`, restart (installed by
  `rootfs/etc/cont-init.d/80-configuration.sh`).
- Add the two `xdg-settings`/`xdg-mime` commands to the custom script
  `/addon_configs/db21ed7f_claude-desktop/claude_desktop.sh` (the image ships no standalone
  terminal).

---

## Problem B — "Your sign-in won't be saved on this device" / recurring re-auth / dispatch shows offline

Symptoms (all three are the same root cause):
- > Install and unlock a system keyring (such as GNOME Keyring), then restart the app.
- Periodic **"For your security, sign in again to keep using Claude."**
- The Claude app's dispatch tab shows this desktop as **not online**, even though the
  process is running — until a fresh sign-in is completed from a computer.

### Root cause (history)
Claude Desktop (Electron) persists its auth token with `safeStorage`, which on Linux
encrypts via the **Secret Service API** (`org.freedesktop.secrets`) provided by
gnome-keyring/kwallet. `--password-store=gnome-libsecret` **forces** that backend.

This was fixed once (v1.17: `gnome-keyring` added to the Dockerfile) and then regressed:
`gnome-keyring` was later removed again because, on first boot, it prompts for a keyring
password and **blocks Claude Desktop from launching at all** — a worse failure than a lost
session. The launch flag was left forcing libsecret, so with no daemon running
`safeStorage.isEncryptionAvailable()` is `false`. Confirmed from a live install's
`~/.config/Claude/logs/main.log`:
```
[safeStorage] isEncryptionAvailable=false ... (backend=gnome_libsecret) — session will not persist; app secrets fall back to plaintext
Electron safeStorage encryption is not available on this system, cannot store allowlist cache
```
The un-persisted session goes stale and then fails the elevated-access OAuth check the
cowork/dispatch bridge needs:
```
permission_error: "Session is not fresh enough to grant elevated access. Sign in again to continue." (session_stale_relogin)
[sessions-bridge] Cowork OAuth stale-session (session_stale_relogin); parking bridge until re-login
```
That "parked" bridge is exactly what the Claude app surfaces as this desktop being
**offline** in the dispatch tab — a fresh sign-in (only completable from a computer, see
Problem A) un-parks it, which is why "open from a computer first" appeared to fix it.

Re-adding gnome-keyring would just reintroduce the original launch-blocking prompt, so it's
a dead end without also solving *that* — hence the shipped fix below avoids keyring
entirely.

### Why v1.35 did not work

v1.35 assumed `--password-store=basic` makes `isEncryptionAvailable()` "always true". It does
not. On a live install the flag was confirmed on the running process:

```
$ tr '\0' ' ' < /proc/2247/cmdline
/usr/lib/claude-desktop/claude-desktop --no-sandbox --disable-dev-shm-usage --password-store=basic
```

and the app still logged, on every launch:

```
[warn] safeStorage not available, tokens will not persist
[warn] Encryption not available, returning empty env vars
[error] Electron safeStorage encryption is not available on this system, cannot store allowlist cache
```

Electron deliberately refuses its own `basic_text` backend unless the **application** opts in
by calling `safeStorage.setUsePlainTextEncryption(true)` before the `ready` event. The symbol
is present in the shipped Electron binary but absent from `resources/app.asar` — Claude Desktop
never calls it. So v1.35 replaced one unavailable backend with another.

Confirmed against a standalone Electron of the same generation, all with
`--password-store=basic`:

| case | `isEncryptionAvailable()` |
| --- | --- |
| no opt-in (= Claude Desktop as shipped) | `false` |
| `setUsePlainTextEncryption(true)` | `true`, `encryptString` works |
| fresh process, decrypting the earlier process's blob | `true`, plaintext recovered |

The third row is the one that matters: it is the restart survival this add-on needs.

### Fix (shipped, v1.37)
1. `claude_desktop/rootfs/defaults/autostart` launches with `--password-store=basic` instead of
   `gnome-libsecret` — Electron's built-in fixed-key store: no daemon, no prompt. Secrets land
   under `$HOME/.config/Claude`, and `HOME=/data/data` is persistent add-on storage, so the
   session survives restarts. `basic` trades away OS-backed at-rest protection: unlike
   `gnome-libsecret`, its encryption key isn't gated by a keyring daemon, so any process able
   to read the persistent `$HOME/.config/Claude` profile can recover the saved credentials.
1b. `claude_desktop/rootfs/etc/cont-init.d/86-claude_safestorage.sh` +
   `claude_desktop/rootfs/usr/local/bin/claude-safestorage-patch.js` (new) supply the opt-in
   the flag depends on, by injecting `safeStorage.setUsePlainTextEncryption(true)` into the
   app's main bundle inside `app.asar`. The injection goes *after* the bundle's leading
   `"use strict";` — a directive prologue only counts as the first statement, so prepending
   ahead of it would drop the main process out of strict mode. The archive is rebuilt (asar
   headers store per-file offsets, so content cannot simply grow in place), preserving
   `unpacked` and symlink entries and recomputing the per-file SHA-256 `integrity` record for
   the single changed file; the result is verified and only then renamed over the original.
   It re-runs every boot after `81-claude_update.sh`, because an apt upgrade of
   `claude-desktop` ships a fresh unpatched `app.asar`; it is marker-guarded, so an unchanged
   app is a no-op, and a failure is logged rather than propagated (an unpatched app still
   runs, it just forgets the sign-in).
2. A passwordless keyring was considered instead (keeps libsecret encryption-at-rest without
   a prompt) and rejected: the keyring DB would live in the same persistent volume as the
   ciphertext it's "protecting," so it adds ~no real confidentiality in this single-user
   self-hosted setup, for more moving parts than `basic`.
3. `claude_desktop/rootfs/etc/cont-init.d/85-openbox_autostart.sh` (new): the app is actually
   launched from the **persistent** `$HOME/.config/openbox/autostart`, which the base image's
   `init-selkies-config` oneshot only seeds from `/defaults/autostart` when the persistent
   copy is *missing*. Editing `/defaults/autostart` alone therefore never reaches an existing
   install's copy. This script re-syncs the persistent copy from the image on every boot (as
   root, before any s6-rc service starts), so this fix — and any future `autostart` change —
   reaches upgrades, not just fresh installs.
4. `gnome-keyring` stays out of the Dockerfile.

### Why v1.37 stopped working

`claude-safestorage-patch.js` only knew how to inject its opt-in *after* a leading
`"use strict"` directive in the app's main bundle, and refused to patch (leaving the app
unpatched and the session un-persisted) if that directive wasn't there. Confirmed live on the
running add-on: Claude Desktop 1.30096.1's main bundle (`.vite/build/index.pre.js`) no longer
opens with a `"use strict"` directive — it now opens directly with a bare IIFE
(`(function(){try{var e=typeof window...`). Upstream's build output changed shape at some point
after v1.37 shipped, the patcher's one injection point stopped existing, and it had been
silently refusing to patch on every boot since — the app's `main.log` kept showing exactly the
same `Encryption not available, returning empty env vars` warning documented above, and the
session went back to not surviving restarts.

`applyPatch()` now falls back to inserting the opt-in as the bundle's first real statement when
no `"use strict"` directive is found, rather than refusing outright. It skips past any leading
BOM, hashbang, or banner comment first (`skipPrologue()`), so a directive hidden behind a
comment is still found and protected instead of being pushed out of the first-statement
position by a naive prepend — Vite/esbuild banners commonly put a license comment ahead of the
directive. A bundle with no directive at all has nothing to protect, so prepending the opt-in
there is unconditionally safe: the injected code is a complete `try{}catch(e){}` statement, and
a statement can never merge with what follows it via ASI the way a bare expression could.

Verified by copying the live production `app.asar` and running the patcher against it directly
(outside the container's boot sequence): the previously-refused bundle now patches
successfully, the marker lands at the front of the main entry, a second run correctly reports
"Already patched" (idempotent), and unit tests cover the bare-IIFE, comment-hidden-directive,
hashbang, and unterminated-comment cases.

### One-time step after upgrading
The previously-stored session is already stale. Complete **one** sign-in from a computer
(mobile still can't finish the OAuth flow itself, per Problem A) — the session then persists
normally and dispatch stays online regardless of which device connects first afterward. This
applies again after the 07308543.1 fix above, since the affected sessions were never persisted
in the first place.

---

## Files this plan touched
- `claude_desktop/rootfs/defaults/autostart` — drop the keyring bootstrap; launch with
  `--password-store=basic`.
- `claude_desktop/rootfs/etc/cont-init.d/85-openbox_autostart.sh` — new in v1.35; syncs the
  persistent autostart from the image on every boot.
- `claude_desktop/rootfs/etc/cont-init.d/86-claude_safestorage.sh` and
  `claude_desktop/rootfs/usr/local/bin/claude-safestorage-patch.js` — new in v1.37; the
  app-side `safeStorage` opt-in that makes `--password-store=basic` actually take effect.
  `claude-safestorage-patch.js` updated again in 07308543.1 to also patch bundles with no
  leading `"use strict"` directive, and to look past leading comments/hashbang when deciding
  whether one is present.
- `claude_desktop/Dockerfile` — corrected stale comment (gnome-keyring is not installed).
- `claude_desktop/CHANGELOG.md` / `config.yaml` — v1.35, then v1.37, then 07308543.1.

Problem A (in-desktop browser for OAuth) remains planned-only; not touched by this change.