Linux/Windows chat client hardcodes api.anthropic.com — no way to point it at a gateway #180

The Anthropic chat client on Linux/Windows hardcodes its endpoint, so the app cannot be used with a self-hosted gateway:

windows/src-tauri/src/claude.rs:14

const ENDPOINT: &str = "https://api.anthropic.com/v1/messages";

The integration set already proves the pattern exists: n8n-url in the keyring is a user-supplied base URL (integrations.rs:611), and every poller takes its host from there. Only the model call is pinned to the public API.

Why it matters. Anything routing model traffic through a gateway — a corporate proxy, a cost/metrics relay, bifrost, litellm — has to leave the island's chat on the public API or not at all, while the rest of the machine uses the relay. Users who never want cloud chat still cannot turn it off at the network layer without losing the integration UI.

Suggested shape, matching the n8n-url precedent: read the base URL from an env var, falling back to the public endpoint when unset or empty.

const ENDPOINT: &str = "https://api.anthropic.com/v1/messages";

fn endpoint() -> String {
    std::env::var("COUCOU_ANTHROPIC_BASE_URL")
        .ok()
        .map(|v| v.trim().trim_end_matches('/').to_string())
        .filter(|v| !v.is_empty())
        .unwrap_or_else(|| ENDPOINT.to_string())
}

and .post(endpoint()) at line 167. Taking the full messages URL (suffix included) avoids guessing whether a given relay mounts the Anthropic surface at /v1/messages or under a provider prefix.

A keyring-backed setting in the existing Settings sidebar would work equally well, and would be more discoverable than an env var — either way, having one of the two is what unblocks gateway users.

Verified on v0.1.3 (Linux port). No telemetry, analytics or updater in the app, so this is the only outbound endpoint that cannot be redirected.

Patch that does exactly the above, if useful: https://github.com/Louis-CFM/coucou/pull (maintainers' call whether to land it as env var, keyring setting, or both).