Models & Providers

Delta’s agents are powered by large language models. You can use models included with your Zed plan, sign in with a subscription you already have, bring your own API keys, or connect a local model.

Hosted models

Paid plans include access to hosted models with usage metered on a token basis. No API keys are required.

In Delta Web, use the model selector to choose which organization supplies your hosted models.

Sign in with a subscription

Four providers connect through an existing subscription instead of an API key:

• ChatGPT: sign in with your ChatGPT Plus or Pro subscription to use OpenAI models.
• Claude: sign in to Claude Agent with your Claude subscription.
• GitHub Copilot: sign in with your GitHub account to use models through Copilot Chat.
• Grok: sign in with an eligible Grok or X subscription to use Grok models.

Connect ChatGPT, GitHub Copilot, or Grok from the desktop app’s settings. These connections sync through your Delta account. Cloud stores the refresh grant and supplies short-lived access tokens to your devices and Cloud Runners. Delta Web uses these subscriptions through a Cloud Runner.

If you connected Copilot before Cloud credential sync was available, sign in again in Settings > LLM Providers. Your existing credential stays in place until Cloud accepts the new connection, but newer clients require the Cloud-backed connection to use Copilot.

Settings > LLM Providers shows the connected account and any account details the provider makes available.

On desktop, ChatGPT Subscription also shows your remaining usage limits and reset times when available. Choose Subscription usage to open ChatGPT’s usage page.

If you reconnect ChatGPT, Copilot, or Grok with a different account, Delta refreshes model access and uses the new account for subsequent requests.

Claude subscriptions

Your Claude subscription works through Claude Agent, not as a model provider for Delta’s agent. Claude Agent runs Anthropic’s agent in your thread in place of Delta’s, with its own sign-in on each computer. Set it up under Agent Connections in Settings > LLM Providers. Claude Agent runs only in the desktop app.

Add an API key

Add your provider’s API key in Settings > LLM Providers. Anthropic, OpenAI, Google AI Studio, OpenRouter, OpenCode Zen, and OpenCode Go support API-key connections. Delta saves provider keys in your account settings and syncs them across devices where the provider is supported.

Google AI Studio API keys are supported only in the desktop app, not in the browser. Requests go to Google; the models do not run on your computer. The available Google AI Studio model list refreshes without requiring a Delta release.

Generate a key in your provider’s dashboard, then paste it into that provider’s API key field. On desktop, if Delta detects an existing key, choose Import existing key beneath the field to copy it into settings. This can import an exported environment variable, or an OpenCode key saved by OpenCode. Importing copies the current value rather than linking to its source: changing the original key later does not change the key in Delta. To rotate or remove it, use Settings > LLM Providers.

For an organization-scoped Anthropic API key on desktop, Delta discovers the key’s workspaces and asks you to select one.

OpenRouter shows a curated set of models that Delta has evaluated for agent work, along with supported models that match Zed-hosted equivalents. You can also add an exact OpenRouter model ID. Delta marks those user-added models as not evaluated, because their agentic capabilities can vary.

On desktop, OpenRouter’s Allowed Providers accepts comma-separated provider slugs to restrict which providers serve your requests. Leave it blank to use your OpenRouter account defaults. Set Provider Order to Highest throughput to prioritize faster providers, or leave it at Default.

Important: Keep API keys out of your repositories. Add them in Settings > LLM Providers, not in project files.

Local models

Connect Delta’s desktop app to a model running on your computer with LM Studio or Ollama. Once you’ve installed models locally and started your server, connect it to Delta:

1. In Delta, open Settings > LLM Providers and choose Add Provider.

2. Enter LM Studio or Ollama as the provider name. Delta automatically fills in the default API URL, selects OpenAI Chat Completions as the API mode, then fetches the available models.

   Server	API URL
   LM Studio	http://localhost:1234/v1
   Ollama	http://localhost:11434/v1

   If you changed your server’s address or port, update the API URL to match.

3. If you configured your server to require an API Key, you can enter it now.

4. Check the listed models and choose Save Provider.

5. Return to your thread. Open the model picker beside the message input and select your local model.

Connect a compatible endpoint

Connect Delta to a local model server, an inference provider, or a gateway that implements an OpenAI- or Anthropic-compatible API.

Custom providers are currently only supported on desktop.

You can also configure a custom provider during desktop onboarding.

Add a provider

1. Open Settings > LLM Providers and choose Add provider.
2. Enter a provider name. Delta uses this name in the model picker. For common providers, Delta tries to prefill the API URL, API mode, and models where possible.
3. Check the prefilled API URL, or enter your provider’s base URL.
4. Enter an API key if your endpoint requires one. You can leave it blank for an endpoint that does not require authentication.
5. Check the API mode, or choose the one your endpoint supports. See Choose the base URL for the available modes and their request paths.
6. Check the models listed under Models, or choose Add model to enter one manually. See Configure models for the model fields and options.
7. If your gateway requires additional HTTP headers, add them under Custom headers.
8. Choose Save provider, then select one of its models in the thread’s model picker.

Choose the base URL

Enter the base URL. Delta adds the path for the selected API mode:

For example, an OpenAI-compatible local server at http://localhost:1234/v1 receives Chat Completions requests at http://localhost:1234/v1/chat/completions.

An Anthropic-compatible gateway at https://models.example.com receives Messages requests at https://models.example.com/v1/messages.

Use the base URL and API mode documented by your provider.

Configure models

Each model needs:

• ID: The exact model identifier accepted by the endpoint.
• Maximum tokens: The model’s context window.
• Maximum output tokens: The maximum number of tokens the model can generate in a single LLM response (text, tool call).

For models that support them, you can also configure Reasoning efforts, Parallel tool calls, and Interleaved reasoning. Leave optional capabilities unset unless your model and endpoint support them.

For OpenAI-compatible endpoints, Reasoning efforts accepts arbitrary effort names, such as adaptive. Of course, it must be accepted by your provider.

For an Anthropic Messages model, Prompt caching offers Disabled, Explicit, and Automatic. Explicit adds cache points with a five-minute lifetime; Automatic requires an endpoint that supports a one-hour cache lifetime. Choose a mode your endpoint supports.

Authentication and custom headers

Enter your API key in the API key field. Delta stores it in your account settings, which sync across devices and use an encrypted local cache.

If your gateway requires additional HTTP headers, add them under Custom headers. Header values are part of your account settings, which sync across devices and use an encrypted local cache. A local endpoint still needs to be reachable from the computer making the request.

Use {thread_id} in a header value to insert the requesting thread’s ID, for example to group requests by conversation in your gateway. We prefill this value for some common providers.

Compatibility and troubleshooting

Delta assumes that custom models support tool calls and images.

If a connection fails:

• Check the request URL. Do not include /chat/completions, /responses, or /v1/messages in the base URL—Delta adds that path.
• Check the API mode. Chat Completions, Responses, and Messages are different protocols.
• Check authentication. Confirm the API key and any headers required by your gateway.
• Check the model ID and limits. Use the identifier and token limits supported by your endpoint.
• For local servers, check that the server is running and reachable from the computer running Delta.

Delta discovers models when you change the connection settings. To retry discovery, click Fetch Models next to Add model. To start over with an empty list, click Clear All Models. Both changes take effect when you save the provider. If discovery finds nothing, add the model manually.

Model roles

Subagents use profiles to choose models. See Subagents for the Worker, Scout, and Reviewer defaults and model configuration for how profile choices and provider-specific preferences interact.

Delta also chooses provider-managed models for its own background work, such as titles and summaries. These are separate from your subagent profile choices. In each provider’s model list in Settings > LLM Providers, a Fast model label marks the model that provider uses for this work. Delta uses it for thread titles, search summaries, turn summaries, and review verdict summaries. On desktop, it also powers Ask in the command palette and the welcome thread. Use Models Shown in Picker in a provider’s settings to choose which models appear when you select a model for a thread.

Use the section’s switch to show or hide all currently listed models at once. Its label is All Shown, Some Hidden, or All Hidden, depending on your selection. Bulk changes apply to the current list; new models appear by default.

Choose a model

You can switch which model a thread uses at any point; the conversation continues with the new model. Use faster models for quick, well-scoped tasks and higher-capability models for open-ended work across many files.

For models that support extended thinking, the model picker has a Thinking toggle and a Thinking Effort menu. Thinking starts enabled in new threads. Higher effort suits harder problems at the cost of speed. Delta remembers the default you choose for each model and applies it to new threads too.

Some models require thinking at certain effort levels. At those levels, the Thinking toggle stays on, and hovering over it explains why. Select a lower effort level to make thinking optional.

Press Ctrl+' in a thread to cycle through the model’s thinking effort levels.

In a shared thread, each participant reaches the thread’s model through their own providers: if you lack the exact provider a collaborator chose, Delta uses the equivalent model from a provider you do have when one is available. Otherwise, the selector marks the recorded model as unavailable. You can keep reading and commenting, but sending asks you to select an available model first.

Automatic fallback

When you run a turn on your computer and a Claude Fable model declines the request, Delta switches the thread to Claude Opus 5 and continues when that fallback is available from the same provider. Delta shows a notification explaining the change. If the fallback is unavailable, Delta shows the refusal.

Context windows and compaction

Every model has a context window: the amount of conversation it can read in one request. Compaction replaces older conversation history with a smaller representation so a long thread can continue.

Click the circular context-usage indicator beside the send controls to open Context Window. There, inspect usage and turn Automatic Compaction on or off when the model supports it. When automatic compaction is on, drag Compact at to set its threshold. The circular indicator shows progress toward that threshold; hover over it for the percentage.

Choose View Context Breakdown to inspect how much of the prompt is used by instructions, messages, thinking, and tools when that information is available. The breakdown shows sizes in bytes and their share of the prompt, not token counts or cost.

Choose Compact Now from Context Window to compact the current thread. On desktop, it is also available from the command palette. You can request compaction while Delta’s agent is responding on your computer; Delta queues the request for the running turn.

Delta explains when manual compaction is unavailable, for example when the agent is running elsewhere, you are editing an earlier message, or there is too little conversation to compact. If compaction fails, use Retry to try again without losing the conversation.

If the provider does not support native compaction, Delta asks the model to write a portable conversation summary. Native compaction behaves differently across providers:

• Supported Anthropic models can compact automatically. Compact Now activates once the conversation is large enough to compact. Anthropic compaction includes a text summary and encrypted metadata for the backend that created it. When you switch to a different provider, Delta keeps the summary and leaves out metadata that the new provider cannot read.
• Supported OpenAI models can compact automatically. OpenAI’s native compaction is encrypted provider state rather than a portable text summary. Delta can reuse compatible state between the OpenAI API and a ChatGPT subscription. Other providers cannot read it, so Delta replays the original conversation when you switch away from OpenAI.

If the provider rejects the conversation as too large after a model switch, Delta asks the selected model to summarize a bounded transcript, then retries the turn once. The summary can still fail or remain too large. If the summary request fails, Delta can offer Retry compaction with followed by the selected model’s name to try again.

If OpenAI rejects encrypted context that the selected model cannot read, Delta offers Switch Back and Retry when it can identify the model that created the context.

What’s next

• Claude Agent covers running Anthropic’s Claude Agent in a thread instead of Delta’s agent.
• Plans & Usage explains how your plan affects model access and spend limits.
• Settings covers Delta’s other configuration options.