Where Ollama Connections Begin: Desktop, CLI, and API Paths

Ollama documents three distinct ways to connect an application:

1. Connect Claude Desktop or ChatGPT Desktop through Ollama’s macOS interface.
2. Configure and start a supported integration with ollama launch.
3. Have a custom program send HTTP requests directly to a local or cloud API.

The practical difference is where the connection begins and where its configuration lives. A ChatGPT Desktop user starts in Ollama’s Apps screen. A Claude Code user can start with the CLI launcher. A developer writing a custom client chooses an API interface and constructs requests.

This comparison is based on documentation, not hands-on testing. It does not assess performance, reliability, security, network exposure, or process lifecycle.

The three paths at a glance

This is a map of the documented interface boundaries, not a description of Ollama’s internal implementation or an exhaustive list of everything that may be possible.

1. Connect a desktop app through Ollama

The desktop path starts inside the Ollama application. Its quickstart documents the following procedure specifically for macOS:

1. Open Ollama and select Apps.
2. Connect Claude or ChatGPT Desktop.
3. Follow the prompts to install or restart the application.
4. Choose Ollama models under Settings → Apps.

In this route, the user makes the connection and selects models through Ollama’s graphical interface rather than writing an HTTP client or invoking an integration by name (Ollama quickstart).

The two documented desktop applications do not have identical boundaries:

• Claude Desktop can use Ollama models in Claude.
• ChatGPT Desktop uses Ollama models in Codex mode. Regular Chat and voice retain their usual ChatGPT connection (Ollama quickstart).

For example, someone who wants to use an Ollama model from ChatGPT Desktop’s Codex mode would choose this Apps workflow. The documentation does not say that connecting ChatGPT Desktop redirects its regular Chat or voice features to Ollama.

The scope of this path is deliberately narrow. The cited quickstart describes the Apps procedure for macOS and names only Claude Desktop and ChatGPT Desktop. It does not provide an equivalent procedure for Windows, Linux, or other desktop applications. That omission is not proof that another arrangement cannot work; it only means such an arrangement is outside this documented workflow.

2. Configure and launch an integration from the CLI

The second path begins with:

ollama launch

Ollama describes this command as an interactive way to configure and start supported external applications that use Ollama models. The current CLI reference lists OpenCode, Claude Code, Codex, VS Code, and Droid as supported integrations (Ollama CLI reference).

This role is distinct from that of ollama run. In the reference, ollama launch handles integrations, whereas ollama run runs a model (Ollama CLI reference).

The documented launcher examples are:

# Select an integration interactively
ollama launch

# Launch a named integration
ollama launch claude

# Select a model while launching Claude
ollama launch claude --model qwen3.5

# Configure Droid without launching it
ollama launch droid --config

These examples establish four specific command forms: interactive selection, launching an integration by name, choosing a model in the shown Claude command, and configuring Droid without starting it (Ollama CLI reference).

A Claude Code user, for example, can ask the launcher to start that named integration and select a model through the documented command. The builder does not have to turn that interaction into a custom HTTP request.

The examples do not establish that every option works identically with all five integrations. Likewise, the supported list defines what the reference currently documents; it does not prove that an unlisted application is incompatible with Ollama.

This boundary depends on the command’s function, not merely on the presence of a terminal. Running a custom API client from a shell does not turn it into an ollama launch integration.

3. Send requests directly from an application

In the direct API path, an application or HTTP client constructs requests and processes the responses. Ollama documents the API as an application-building interface for both cloud and local models (Ollama API introduction).

The application chooses among these documented base-URL forms:

These are options within the direct API route, not three more top-level connection paths. In the Anthropic configuration, the client appends /v1/messages to the selected base URL; the table row is therefore a client configuration value, not a native Ollama API endpoint (Ollama API introduction).

For example, a developer writing a custom chat client could send a native Ollama request to one of these endpoints:

https://ollama.com/api/chat
http://localhost:11434/api/chat

The developer owns the base URL, headers, request body, and response handling. That control is the defining feature of this path.

Direct cloud access requires an API key but does not require an Ollama installation. Running a model on the computer requires Ollama and a local model. Local requests do not require an API key, while using cloud models through a local Ollama server requires signing in (Ollama API introduction).

The absence of an API-key requirement for local requests is only a credential statement. It does not establish that a particular deployment is private, secure, or inaccessible over a network. Nor do the listed base URLs establish feature parity among the native, OpenAI-compatible, and Anthropic-client interfaces.

A concise selection rule

Choose the macOS Apps workflow for the documented Claude Desktop or ChatGPT Desktop connection. Choose ollama launch when Ollama’s CLI should configure and start one of its listed integrations. Choose the direct API when your own application should create requests for a local server or Ollama Cloud.

None of these routes is inherently ranked above the others here. They assign setup and request handling to different interfaces—and that boundary is the useful basis for choosing among them.

Sources and further reading

• https://docs.ollama.com/quickstart
• https://docs.ollama.com/cli
• https://docs.ollama.com/api/introduction