Skip to main content
The Claude Desktop app hosts the Code and Cowork tabs. ValarCode routes them through Valar with valar claude-desktop, which writes a Valar profile the app reads at startup. Both tabs keep speaking the Anthropic Messages API and you keep picking Claude models; the gateway maps whatever they send to the target your cohort resolves to. See Set up ValarCode for the shared setup and the client-id model.
valar claude-desktop is separate from valar claude, which routes the Claude Code CLI. The app keeps its routing in its own config store, so the two are enabled, disabled and reported apart. This is a macOS-only integration.

​
Prerequisites

  • macOS with the Claude desktop app installed (the app that hosts the Code and Cowork tabs).
  • A Valar coding key (vlrcode_…) and the valar CLI (install).

​
Enable routing

Run once per machine:
The app reads its config only at startup. If it is running, valar offers to quit and relaunch it for you — pass --force to skip that prompt and restart it right away, at the cost of interrupting any session open in the app. Decline, and the change applies on the app’s next launch. Check the state any time:
The status output reports whether the app is installed and whether it is currently routed through Valar. Routing covers the Code and Cowork tabs. The Chat tab is opt-in, per run:
The choice is not remembered — pass --with-chat on every on that should keep the tab, including the re-enable an upgrade performs. Omitting it removes the Chat tab, which Code and Cowork never do.
The Chat tab starts a separate history: chats made through Valar stay here, and your claude.ai chats stay on your account. Its traffic also runs through the gateway and bills as this harness.

​
What gets written

The CLI writes a profile named Valar into the desktop app’s third-party config store under ~/Library/Application Support/Claude-3p/configLibrary/:
  • _meta.json selects the Valar profile and preserves any profiles you already had.
  • <uuid>.json holds the profile body:
In detail, enabling:
  • Points the app’s inference at Valar (inferenceProvider: "gateway", base URL written as the host without a /v1 suffix). Both tabs route over the Anthropic Messages surface.
  • Uses the bare coding key as a static bearer (inferenceCredentialKind: "static").
  • Pins the model picker with inferenceModels and sets modelDiscoveryEnabled: false, so the picker shows the Valar models rather than filling from the raw catalog. The first entry is the chat default (for example Sonnet), since Cowork is a chat surface.
  • Carries attribution in inferenceCustomHeaders. This is how Cowork’s isolated sessions stay attributed: they never see the ANTHROPIC_CUSTOM_HEADERS env that the Code tab and Claude Code use, so the client id and harness ride on the profile instead.
Before the first change, the app’s existing config is snapshotted to ~/.valar/claude/desktop-backup.json.
The client id defaults to your OS username, so the header reads X-Valar-Client-Id: jdoe. Use --user-id RANDOM for an opaque vc_… id instead. See per-engineer attribution.

​
Route a single instance

To run a Valar-routed Claude Desktop beside your main app, without touching its config:
This launches a separate Claude Desktop instance pinned to its own data dir (~/.valar/claude/desktop-session) and routed through Valar. Your main app, its login and its config stay untouched, and valar claude-desktop status stays off. One single-session instance runs at a time. If a previous one is still open, pass --force to replace it:

​
Turn off routing

This restores the pre-enable snapshot byte-for-byte. If the CLI created the whole Claude-3p config tree, it removes it. As with on, valar offers to restart the app for you, and --force skips the prompt.

​
Next steps

Claude Code

The Claude Code CLI, routed separately, and how it edits ~/.claude/settings.json.

Model routing

How cohorts and the split decide which model serves each request.