Zed with Nebius Token Factory

Use Nebius Token Factory for Zed-owned AI features through Zed's existing OpenAI-compatible provider. No Zed extension, native provider, or Token Factory adapter is required.

This recipe configures moonshotai/Kimi-K3 over the OpenAI Chat Completions protocol. It applies to the Zed Agent, Inline Assistant, commit-message generation, thread summaries, and similar model-backed Zed features. It does not configure External Agents, Terminal Threads, or Edit Prediction; those are separate Zed surfaces with their own model access and authentication.

Prerequisites

• A current Zed installation
• A Nebius Token Factory API key
• Python 3.11 or newer for the optional offline validator

Configure the provider

Open Zed's Agent Settings from the command palette with agent: open settings. Under LLM Providers, choose Add Provider, then select an OpenAI-compatible provider and enter:

The provider ID is significant. Zed converts it to upper snake case and appends _API_KEY; nebius therefore reads NEBIUS_API_KEY.

Enter the Token Factory key in Zed's provider UI to store it in the local system keychain. Alternatively, start Zed from an environment that contains the key:

export NEBIUS_API_KEY="your-token-factory-api-key"
zed .

Do not put the key in settings.json.

Merge the maintained settings

The UI writes equivalent settings. For a reviewable manual setup, open your Zed settings file with zed: open settings file and merge the contents of settings.json into your existing top-level JSON object. Do not replace unrelated editor settings.

The important protocol lock is:

"capabilities": {
  "chat_completions": true,
  "interleaved_reasoning": true,
  "max_tokens_parameter": false
}

With chat_completions: true, current Zed appends /chat/completions to the configured API URL. Setting it to false switches Zed to /responses, which is outside this recipe. interleaved_reasoning allows reasoning-capable chat models to round-trip the dedicated reasoning_content field. Keeping max_tokens_parameter: false uses max_completion_tokens, which Token Factory's Chat Completions API accepts.

Select and use the model

The settings fragment makes Kimi K3 the default Zed model. You can also select Nebius Token Factory · Kimi K3 from the model picker in the Agent Panel.

Start with a low-risk prompt and approve tool actions deliberately. Provider configuration does not change Zed's tool-permission policy.

Validate offline

The validator parses the settings without launching Zed or contacting Token Factory. It checks the exact schema subset, base URL, model, context, provider-to-key mapping, default-model reference, secret exclusion, and Chat Completions capability.

python validate_settings.py
python -m venv .venv
.venv/bin/pip install -r requirements-test.txt
.venv/bin/pytest -q

Limitations

• This recipe covers Zed-owned AI features only. External Agents and Terminal Threads are separate processes and do not inherit this provider.
• Edit Prediction uses a separate configuration and the legacy OpenAI /v1/completions format. Kimi K3 is not configured here as a fill-in-the-middle edit-prediction model.
• The current model is pinned because its 1M context, vision, tools, and reasoning declarations are model-specific. When changing models, update and validate all of those fields together.
• parallel_tool_calls and prompt_cache_key remain off because this recipe does not claim those optional Zed request features for Token Factory.
• Zed's OpenAI-compatible provider rate-limits itself to four concurrent requests.
• API access is initialized in the local Zed app. For SSH and dev-container projects, keychain credentials and provider environment variables come from the local Zed process, not the remote shell.

Troubleshooting

Zed asks for an API key

Enter it through the nebius provider UI, or make sure NEBIUS_API_KEY is in the environment of the Zed process and restart Zed. A key exported only in an embedded or remote terminal does not automatically reach the local app.

401 or 403

Create or rotate the key in Token Factory, reset the provider credential in Zed, and enter the new value. Confirm the API URL still ends in /v1.

404 or request reaches /responses

Keep the base URL at https://api.tokenfactory.nebius.com/v1 and capabilities.chat_completions set to true. Do not append /chat/completions yourself; Zed appends it. A false capability sends the request to the Responses endpoint instead.

Model not found

Confirm the exact current model ID in the authenticated Token Factory model catalog. If Kimi K3 has been retired, update the model ID, context window, capabilities, default-model reference, and validator together instead of changing only the name.

Tools or images fail

Check that the model remains advertised with those capabilities in Token Factory. If the selected replacement model lacks either feature, turn the corresponding Zed capability off so the UI does not offer unsupported inputs or tools.

Documentation gap audited

Token Factory's public integrations overview lists Zed under Coding Assistants, but as of this audit the label has no linked setup page or configuration details. This first-party cookbook recipe fills that onboarding gap without adding or duplicating provider code in Zed.