Use NVIDIA models with Claude Code or Claude Desktop
RouteMyAgent provides an Anthropic-compatible Messages gateway for your rented NVIDIA model. Your coding client keeps its local workflow, while model requests go through the rented route using the NVIDIA key you attach in the portal.
Claude Code CLI and Claude Desktop’s Code tab have different setup paths. Choose the matching client in the portal. A CLI environment variable is not proof that the graphical Desktop application uses that gateway.
Before connecting
Working to a budget? The AI coding cost guide compares hourly routing with direct API access, self-hosting and access you already have. Provider and client costs remain separate.
- Have an authorized NVIDIA account and key ready.
- Install your chosen Claude client and check that its version exposes the required custom-gateway configuration.
- Keep your existing account, permissions, MCP configuration and normal launcher. Client account requirements still apply.
- Choose the rental’s model. Subagents and model aliases must resolve to that same model, not a different tier.
The BYOK guide explains the two credentials and how prepaid time works. The free promotion covers 30 minutes of router access, not NVIDIA credits or a Claude subscription.
Claude Code CLI: a separate rental launcher
- In the active rental, select Claude Code CLI and copy the generated setup prompt.
- Give the prompt to a local coding agent, or have a browser chat prepare runnable setup files for your operating system. Do not paste either credential into that conversation.
- The setup creates a dedicated directory, Start and Undo launchers, a README and a change receipt. It inspects existing configuration instead of overwriting your providers, tools or authentication.
- The Start launcher reads the router token privately and sets gateway configuration for its child process. The setup must detect conflicting saved settings instead of silently routing to the wrong service.
- Verify a small response and a harmless tool call with its result, then use the rental launcher for the session.
The hosted gateway base is https://routemyagent.dev/claude, with Messages at /claude/v1/messages. Copy the exact base and reversible model identifier from your rental; the compatibility alias still selects the same NVIDIA model. Use the router token in the client, not your NVIDIA key.
Claude Desktop: use the actual Gateway configuration
Select Claude Desktop · Code tab in the portal. Use the installed Desktop version’s third-party inference Gateway configuration and the generated Desktop instructions. Preserve the existing connection profile so you can switch back. Enter credentials through the supported private credential mechanism, never into a shared setup prompt.
After applying the Desktop configuration, start a local Code session, select the rented model and verify that the gateway receives the request. Run a small file or tool task before a longer session. Configuring a local Gateway does not reroute unrelated hosted browser chats.
Subagents and parallel work
The router can carry supported tool calls used by client-managed subagents. Desktop background tasks and parallel local sessions remain features of the installed client. Keep their model selection on the rented model. The current hosted service admits up to four simultaneous requests per rental, counting queued requests and open streams. A fifth request receives HTTP 429 with Retry-After: 1; follow the returned delay rather than repeatedly resubmitting.
Requests sharing a NVIDIA key share the router’s 10 requests per minute pacing, with at least 6.05 seconds between request starts. NVIDIA can impose additional limits. All agents share the remaining usage time (or a legacy rental deadline); overlapping requests count once per account, and more agents do not create additional provider quota or NVIDIA capacity.
Model behavior, tool quality and client features can vary. RouteMyAgent implements compatibility for NVIDIA models; this is not Anthropic certification or a promise that every Claude-specific feature is available. Provider-hosted state and signed reasoning are not fabricated by the adapter.
Return to your normal setup
For CLI use, close the rental child and open your normal launcher. Preview the scoped Undo script before removing unchanged setup-owned files. For Desktop, switch back to the saved original connection profile and review the generated change receipt. Undo does not require an active rental and does not refund purchased time. New usage time pauses when no model request is running; older wall-clock rentals retain their deadline. If an older setup edited shared settings, review the exact changes rather than restoring a whole old configuration over newer work.
Public undo instructions remain available after expiry. For connection errors, check the rental, attached key and exact model first; do not repeatedly submit a request whose completion is unknown.
Client documentation
For version-specific details, see the official Claude Desktop Gateway guide, in-app configuration guide and Desktop Code documentation.