Zoo Code + JoinGonka Gateway: Migrating from Roo Code to Gonka models

Step 1: install the extension. In VS Code, open the extensions panel (Ctrl+Shift+X, on macOS Cmd+Shift+X), search for "Zoo Code" and install the extension from the publisher Zoo Code Organization. The same in one command:

code --install-extension ZooCodeOrganization.zoo-code

You need VS Code 1.100 or newer — as stated in the extension manifest; for VSCodium, Windsurf and other editors without the Marketplace, the extension is published on Open VSX. Besides the stable channel (3.82.2 from September 18, 2026), there's a preview 3.83.x — nightly builds of the main branch; for work, use the stable one. After installation, a Zoo Code icon will appear in the activity bar.

Step 2: get a key. Register at gate.joingonka.ai/register, confirm your address and create a key with the prefix jg- in the "API keys" section. It's convenient to create a separate key for Zoo Code: its usage will show up as a separate line in the statistics.

Step 3: run the installer.

npx @joingonka/setup --tool zoo --model deepseek

The installer will ask for the key — it isn't passed in the command-line arguments so it doesn't end up in the shell history. It has nothing to write: provider profiles along with Zoo Code keys are kept in VS Code's secure secrets storage, not in a file. That's why the installer prints ready-to-use values for the form and the real limits of the selected model:

Zoo Code is configured in the VS Code extension UI (no file is written).
Open the Zoo Code settings (gear icon) → API Provider, then enter:

  API Provider: OpenAI Compatible
  Base URL:     https://gate.joingonka.ai/v1
  API Key:      jg-your-key
  Model:        deepseek-ai/DeepSeek-V4-Flash-0731

Then set the real limits of this model under Model Configuration
(the defaults do not fit any Gonka model):

  Context Window Size: 380000
  Max Output Tokens:   32768
  Image Support:       off

Zoo Code relies on native tool calling only — all our Gonka models support it.

Next comes the result of a live request to the gateway — on success, the line ✓ Verified: the gateway accepted the key, base URL and model. The key, address and model are verified immediately, not on the first task. The key in the printed block is shown in full so it's easy to copy — don't publish this output.

Without the --model flag, the installer will substitute MiniMax M2.7: where the model is chosen manually, it suggests the model with the highest capacity in the network. For Zoo Code, we recommend starting with DeepSeek V4 Flash: in our test, it completed the turn after the tool result on the first request — details in the verification section; the flag also accepts the abbreviations glm (GLM-5.3 Flash) and minimax. Without the --tool flag, the installer will show a menu of 25 tools.

One form detail: there's no separate block titled Model Configuration in Zoo Code 3.82.2. The limit fields are in the same form below, under the explanation "Configure the capabilities and pricing for your custom OpenAI-compatible model", and Max Output Tokens comes before Context Window Size.

Upon the first launch, Zoo Code displays a greeting with a Choose your provider button: it leads to a form where OpenRouter is pre-selected — change the provider to OpenAI Compatible and click Finish at the end. Later, the same form can be opened via the gear icon in the Zoo Code panel header, under the Providers section; the Save button applies changes.

Network model limits:

You don't need to touch anything else: leave Enable streaming and Include max output tokens enabled, keep Enable R1 model parameters and Use Azure disabled (these are flags for the R1 family and Azure OpenAI), and leave the Custom Headers and Extra Body fields empty.

Limits are stored in the profile, not tied to the model: if you change the Model in the same profile, the window and output limit will remain the same. Therefore, it is more convenient to create a separate profile for each model — this is covered in the next section.

Zoo Code evolved from the Roo Code codebase, so the migration takes a few minutes and project files remain untouched. The official migration guide covers exporting and importing settings; everything else is transferred via a separate button or read from the old paths:

A Roo Code profile that was already pointed at JoinGonka Gateway will migrate with the URL, key, model, and limits — just verify that the model is currently available in the network: the list is provided by GET https://gate.joingonka.ai/v1/models.

Modes and Profiles. There are five built-in modes: Code, Debug, Architect (edits only Markdown files), Ask (answers without edits), and Orchestrator (assigns subtasks to other modes). They can be switched via the list next to the input field, by typing a command with the mode name at the beginning of the message (e.g., /architect), or by using the shortcut Ctrl+. (Cmd+. on macOS). Each mode remembers the profile you last used with it, and Zoo Code automatically switches profiles when you change the mode.

This creates a convenient scheme: one profile per model, tied to specific modes. In Settings → Providers, click "+" next to the Configuration Profile field, name the profile (e.g., Gonka DeepSeek), and fill out the form; repeat for GLM-5.3 Flash and MiniMax M2.7 — the URL and key are the same for all. Then, select the profile in the profile list next to the input field in each mode, and that mode will remember it. Conversely, the lock icon in this list pins one profile for all modes in the workspace folder.

Reasoning Effort. The Enable Reasoning Effort checkbox in the form opens a list: Low, Medium, High, Extra High, and Max; while it is disabled, the level is not included in the request. For GLM-5.3 Flash, the toggle is essentially binary: Low turns off reasoning, any other level leaves it fully enabled — it is convenient to keep a separate GLM profile with Low for quick questions. For details, see the GLM-5.3 Flash overview. For DeepSeek V4 Flash and MiniMax M2.7, the gateway does not rewrite the scale: how the level works depends on the model itself.

Open a folder with a small file containing an obvious error, for example calc.py with an addition function that subtracts, and in Code mode, type: "Read calc.py and say in one sentence if it has an error". Zoo Code will ask for permission to read the file — "Zoo wants to read this file", with Approve and Deny buttons if reading was not allowed in advance via Auto-Approve. Then the model will call a completion tool, and the task will end with a Task Completed block with the actual answer; reasoning, if any, is shown in a separate Thinking block. The context usage percentage in the collapsed task header is calculated based on the Context Window Size — another reason to set it correctly.

We reproduced this workflow by issuing a request using the Zoo Code pattern — flow, tools in a strict schema, result of read_file in history. DeepSeek V4 Flash responded with a native attempt_completion call with the correct conclusion: the function subtracts instead of adds. The second half of the check is the gateway dashboard: in the "Usage" section, the request will appear broken down by model, and the key's last used time will be updated.

Without Auto-Approve, an error appears as a API Request Failed block with Retry and Start New Task buttons, accompanied by a Provider Error header with a code and a countdown to retry. The gateway's response text is visible in the message or its details:

An agent spends tokens differently than a chat. In every request, Zoo Code includes a system prompt with mode rules, tool descriptions, and information about the working folder — that's thousands of tokens even before your question — and a task usually takes several turns. Therefore, the price per token is critical here.

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

For comparison — how you can pay for models in Zoo Code in general:

With Input Price and Output Price filled in, the cost of the task is visible in its header, but only down to the cent — short tasks on network models will show $0.00 there. The exact consumption and balance are in the dashboard, in the "Usage" and "Billing" sections.

Approvals and autonomy. By default, Auto-Approve is disabled, and Zoo Code asks for permission for reading, writing, and executing commands; for long tasks, enable the necessary categories in the Auto-approve menu near the input field. For commands, there is a Destructive Command Guard (Settings → Auto-Approve): the extension downloads its executable file, after which the commands it approves are executed automatically, while blocked ones wait for confirmation. With Auto-Approve, Zoo Code also retries failed requests automatically: the pause starts at 5 seconds and doubles with each attempt, but does not exceed 10 minutes.

Streaming response. Do not turn off Enable streaming: in stream mode, the gateway delivers text as it is generated and limits the response only by the model's ceiling, whereas a request without streaming and without an explicit limit is truncated at 1500 response tokens (for GLM-5.3 Flash — at 3000).

Code Indexing. Semantic search through the project (Codebase Indexing) is configured separately from the model, via an embedding provider; there is a local option for it, Semble - Local.

Images. The network models are text-based. For screenshots and mockups, create a separate profile with a model that works with images and switch to it for the duration of such a task.

Privacy. Zoo Code sends error and usage data to the developers, linked to an installation ID — without code and prompts; this can be disabled via the Allow error and usage reporting toggle in Settings → About Zoo Code. For its part, the gateway does not store the content of prompts and responses: only request and token counters are visible in the dashboard.