Warp + JoinGonka Gateway — terminal agent on your own endpoint

JoinGonka Key. Register on the gateway, open the "API Keys" section in your dashboard, and click "Create key". The key starts with jg- and is shown only once—save it immediately. It is convenient to create a separate key for Warp: this way, usage in the "Usage" section for the terminal agent will be visible as a separate line item, rather than mixed with other tools.

Warp Plan. Your own endpoint is available not just on paid plans—conditions depend on the team size:

Public HTTPS Address. This is a Warp requirement that is easy to miss: requests to your endpoint are sent not by the application on your machine, but by Warp's servers, so the address must be accessible from the internet. The form rejects localhost, 127.0.0.1, and private network addresses, and only the https:// scheme is allowed. JoinGonka Gateway is a public HTTPS service, so tunnels like ngrok, which Warp documentation suggests for local models, are not needed here.

Latest application version. Your own endpoint was included in the stable build starting from the May 20, 2026 release, and the API schema selection in the form appeared in the July 31, 2026 release. If the API schema field is missing in the form, update Warp.

Everything is configured in the app interface — no need to edit files by hand.

1. Open Settings and type inference endpoint in the settings search — Warp will jump to the section with provider keys and your own endpoints (it lives on the Warp Agent settings page).
2. Click Add custom endpoint — a form will open.
3. Fill in the fields with the values from the table and click Add endpoint.

The + Add model button adds another row. Add all the network's models at once so you can switch between them later without going back into settings: deepseek-ai/DeepSeek-V4-Flash-0731, MiniMaxAI/MiniMax-M2.7, and zai-org/GLM-5.3-Flash. The current list is always available at GET https://gate.joingonka.ai/v1/models. Enter the IDs character by character: the form only validates the URL format and that the fields are filled in — it doesn't send a test request on save, so a typo in the model name will only surface later in a conversation with the agent.

Once saved, the models appear in the agent's model switcher. If the default model was still one of Warp's models, the app may offer to change it — accept and pick your endpoint's model. One important detail: the Auto options in the switcher always run on Warp's infrastructure and burn its credits, even when your own endpoint is configured. For a request to reach the gateway, a specific model from your list must be selected in the switcher.

Choose your endpoint's model in the switcher and give the agent a short task, for example: "Show me the five largest files in the current directory and explain the command." The agent should propose a command, run it with your permission, and comment on the output.

Now open the gateway dashboard and go to the "Usage" section. Under "By keys", the key you created for Warp will show requests and the time of the last call; under "By models", the selected model. If the counters are growing, inference is going through JoinGonka Gateway, not burning Warp credits. For its part, Warp shows token usage right in the chat with the agent — under the short names you set in the alias field.

If something goes wrong, the reason is usually clear from the response:

The first response in a session may take a few seconds: the agent sends a large system prompt describing the tools, and the request first passes through Warp's servers and then a network node. After that, the conversation moves faster.

All network models have the same price, so your choice is about behavior, not budget. The terminal agent reads a lot (command output, logs, files) and calls tools frequently; all models in the table support native tool calling.

Practical scheme: DeepSeek V4 Flash is the primary model, MiniMax M2.7 is for small tasks, GLM-5.3 Flash is for when the task is logic-intensive rather than volume-intensive. Switching takes a second because all models are already configured in the endpoint, and short aliases prevent confusion with long identifiers. More details on the most popular one in the DeepSeek V4 Flash overview.

The terminal agent has higher token consumption than a chat: for every request, Warp adds system instructions, dialog context, and tool descriptions, and one task consists of multiple turns. Therefore, the price per token matters more here than in casual chatting.

Via JoinGonka Gateway, tokens cost $0.0069 per million for input and $0.021 per million for output—the price is the same for all network models and is pulled onto this page from a live source. Price scale as of September 2026:

For comparison, three ways to pay for the Warp agent:

Warp itself does not charge credits for individuals and teams of up to 10 people when using your own endpoint—you only pay for tokens at the gateway. The balance is topped up in your dashboard, where you can also see your remaining balance and daily usage by model.

For most tools, a custom endpoint means “requests go directly from my machine to the provider.” Warp is different: the executing part of the agent runs on Warp servers, and a custom endpoint only changes the final destination and the key. According to the Warp documentation, the request path is as follows:

1. The application retrieves the endpoint address and key from a secure storage on your device and sends them to Warp servers along with your prompt.
2. The agent on the Warp side assembles the complete request—system instructions, chat context, tools—and calls your endpoint with that key.
3. The response is returned to the application as a stream through the same servers.

Practical implications:

• Key. It is stored only on your device. It passes through Warp servers with every request, but according to Warp, it is not saved there: it is used on the fly and discarded.
• Prompts and responses. They pass in transit through the Warp backend. Warp states that it does not use this content for training, and storage and analytics are subject to your account's privacy and telemetry settings. It cannot extend the ZDR guarantee, which applies to Warp's own models, to your custom endpoint: what the provider stores is determined by the provider.
• Gateway side. The JoinGonka Gateway does not store conversations: prompts and responses do not remain on the gateway after the response, and only request and token counters are visible in the dashboard.
• What remains on Warp infrastructure. Auto options, custom model routers, Codebase Context, Active AI suggestions, and cloud agents—a custom endpoint does not affect these.

If the path through third-party servers is unacceptable for your tasks, use an agent that accesses the endpoint directly from your machine, such as Codex CLI or any tool from the API quickstart guide. In its announcement, Warp itself promises a client-side option: a lightweight agent module in Rust in an open repository that will connect to local models without routing through Warp servers, as well as support for the Agent Client Protocol. And as of the September 2, 2026 release, endpoint descriptions are stored in a configuration file in a format common to both the application and the Warp Agent CLI; keys do not end up in this file and remain in secure storage.