Connect Codex or OpenCode to your NVIDIA rental
RouteMyAgent gives these coding clients an OpenAI-compatible API using your own NVIDIA account. The model behind the route is the selected NVIDIA model, not an OpenAI model. Your local client still manages files, tool execution, MCP connections and its permission system.
Choose the matching protocol
Before choosing a rental, compare the total cost of AI coding, including your provider access and any client requirements. Routing fees are separate from both.
Codex CLI uses the Responses route: the provider base is https://routemyagent.dev/v1 and requests go to /v1/responses. The rental setup supplies the exact model, provider definition, context metadata and launch arguments for a separate process.
OpenCode uses Chat Completions: the same /v1 provider base leads to /v1/chat/completions. OpenCode v1 and v2 have different configuration schemas. Select the matching version in the portal; do not combine one version’s field names with the other’s.
Use the portal’s generated configuration for your active rental. Guessing a model ID, reusing an expired rental token or adding an extra /v1 to a route can produce errors before a model receives the request.
Codex: configure one rental process
- Check the installed CLI version and supported arguments. Choose Codex CLI in the portal.
- Use the setup prompt to create a dedicated Start launcher, Undo launcher, README and receipt. Keep normal providers, auth, tools and history in place.
- The launcher supplies the rented model and provider through process-scoped CLI overrides. It obtains
AGENTNET_ROUTER_TOKENprivately for the child process. Do not replace your normal global config or place a key in a prompt. - Use the provided context limits. A model’s advertised maximum is not proof of answer quality at every length, and output allowance differs from total context.
- Verify text, then a harmless tool call and its result through the actual installed CLI.
Namespaced function tools are translated and restored so client-managed tool identities remain intact. This stateless route requires complete conversation history. It does not provide provider-stored conversations or encrypted reasoning state. Cache-key hints do not guarantee caching, and client metadata is not forwarded to NVIDIA.
Provider-hosted built-in web search is unavailable on this adapter. The generated rental configuration disables that unsupported Codex search mode; separately configured local or MCP tools remain available. Client-managed multi-agent tools can still use the rented model within the same provider quota and remaining usage time (or legacy deadline).
OpenCode: match its installed schema
Choose OpenCode v1 or OpenCode v2 only after checking your installed version. The generated reference sets the selected model’s context and output limits, supported input modes and reasoning mapping. Treat missing capabilities as unknown rather than enabling them from the model name alone.
The setup uses a separate process-scoped configuration path where that OpenCode version supports it. It checks how existing settings merge and reports conflicts. A CLI configuration does not automatically configure every Desktop build: verify how that version launches and inherits its environment.
What the connection test should prove
First check the exact endpoint, active rental and allowed model. Then ask for a short response. Next run a harmless local tool, return its result to the model and confirm the final answer uses it. This checks the protocol’s multi-turn path as well as basic text. The portal’s optional Test connection button only checks the hosted route; it cannot prove your local client is configured correctly.
On 401, inspect the current rental token. On 429, honor Retry-After and reduce parallel work. On an interrupted response or gateway timeout, check completion before retrying; the provider may already have processed the request. Never switch models silently to conceal a configuration error.
Keep setup easy to reverse
Run the generated Undo preview before activation. It should list only the files that setup owns, with no network or active rental required. When finished, close the rental child and open your normal coding client. The scoped Undo script can remove unchanged rental files while preserving later edits. Revoking a router token is a separate portal action; neither action refunds purchased time. New usage time pauses when no model request is running; older wall-clock rentals retain their deadline.
A free browser chat can prepare these files and commands, but needs local execution access to apply them. A setup prompt cannot grant a client subscription or change the backend of an unrelated hosted chat.