RouteMyAgentConnect MCP ↗

LOCAL CONNECTOR PREVIEW · OCTOBER 26, 2026

Choose a Provider for Your Local Connector

Windows public preview. Download the unsigned AgentNet Windows package from setup. Verify its SHA-256 before opening it. Installed-client/provider operation and macOS/Linux/mobile support are not verified by this download.

Browse NVIDIA and Cerebras models and exact IDs, then prepare a local connection for your coding client or MCP host.

Provider and model directory ↗

  1. Choose the connection. Select your provider, exact model and installed Claude Code, Codex or Hermes client, then choose agent or MCP mode. Your provider account stays yours.
  2. Prepare managed setup. Sign in and explicitly start the no-card trial or use an active subscription to retrieve a credential-free local connection file. Account access is separate from provider authentication.
  3. Import locally. Download the unsigned Windows preview package, verify its SHA-256 and import the connection in AgentNet. Enter the provider key only in its native local UI and select its exact OS-stored credential reference. Never paste a provider key into the RMA website, chat or setup JSON.
  4. Verify your own client. Review the generated configuration, preserve the previous settings and scoped Undo, then authorize one small direct-to-provider request. No automatic retry or provider/model substitution is authorized by setup.

Catalog facts are not execution proof

Task labels, context/output limits and advertised API types are metadata. An upstream embeddings, ranking, image, audio, video or OCR path does not add that operation to a client adapter or prove the local connector supports it. Unknown limits stay unknown. Use the exact model and the installed connector/client’s supported wire format.

Provider-direct execution

In the local connector preview, AgentNet reads the selected credential from the operating system’s credential store and contacts the provider directly. RMA manages account access, configuration and compatibility metadata; provider keys and inference content do not pass through RMA on that local path.

Public RMA provider metadata is discovery/configuration data, not a hosted inference URL. Your generated local configuration identifies the actual connector address; no provider key belongs in a website form, URL or configuration download.

Provider capabilities and billing

Use your own authorized provider account. Provider access, inference charges, quotas, model rights and data policies are separate. Catalog metadata and generated recipes are not proof of a successful installed-client or provider request.

Maintain full tool-call IDs/results and any provider-required reasoning history. A compatibility recipe does not make signed thinking, encrypted reasoning or every native provider feature representable by every client.

Models on your own computer

A local model API is not automatically an MCP server. Use an explicitly supported local runtime/MCP configuration in your connector and verify it locally. Do not add an unverified bridge or switch providers to conceal an unsupported operation.

Earlier hosted connections and balances

The earlier hosted path sent provider keys and model content through RMA. Do not reuse its provider-key headers or hosted inference URLs for this preview. Earlier prepaid balances, receipts, payment identities and applicable rights remain recorded separately, without automatic conversion or forfeiture. Contact support about an unused balance or an uncertain earlier payment; do not pay again to recover a record.