Ollama Cloud for GitHub Copilot Chat

Use hosted Ollama Cloud models directly from the GitHub Copilot Chat model picker in Visual Studio Code—without installing Ollama or running a local model server.

This native VS Code LanguageModelChatProvider validates an Ollama API key, discovers hosted models, and streams responses directly from ollama.com into Copilot Chat without a local proxy.

Highlights

• Direct Ollama Cloud integration with no local Ollama installation
• Native entry keys owned by VS Code Secret Storage, scoped by a one-way key fingerprint
• Live discovery with six-hour persisted models.dev enrichment
• Streaming text, separate thinking parts, image inputs, and agent-mode tool calls
• Verified model-specific thinking controls in the model picker
• Published per-model input, cached-input, and output pricing metadata
• Optional Ollama Cloud Web Search tool for agent mode
• Resilient native and fallback token accounting
• Status-bar five-hour and weekly subscription usage

Quick start

1. Install Ollama Cloud for GitHub Copilot Chat. You need VS Code 1.125 or newer, GitHub Copilot Chat, and Ollama Cloud access.
2. Create a key in Ollama API keys.
3. Open Manage Language Models, choose Add Models, select Ollama Cloud, name the entry, and paste the key.
4. Repeat the previous step for any additional account or API key, then enable the models you want in Copilot Chat.

Composer controls override workspace defaults; ordered thinking controls default to High and verified binary controls default On. Click the Ollama status-bar item to inspect account utilization, per-model requests, and tokens observed by this extension. Account utilization and local token totals remain separate because Ollama limits are GPU/time based.

Management commands and inline suggestions use the active key (the most recently provisioned entry in Manage Language Models).

Images returned by tools are sent as native image attachments after the corresponding tool results; their base64 bytes are never inserted into prompt text. Context token counts remain estimates. If a request exceeds the context window, run /compact in Copilot Chat, reduce attachments or tool output, or start a new chat.

Tool calls must match the current request's advertised tools. Whitespace and a functions. namespace are repaired only when they resolve to an exact advertised name. A rejected call prevents the whole batch from being forwarded; its error includes structural diagnostics without exposing tool names or arguments.

Documentation

• Setup, commands, settings, and troubleshooting
• Models, thinking, and accounting
• API key and security model
• Development and releases

Related projects

• Codex Bridge for Copilot Chat
• Grok for GitHub Copilot Chat
• OpenCode for Copilot Chat
• Poolside for GitHub Copilot Chat

Unofficial project; not affiliated with Ollama, GitHub, or Microsoft. Ollama account limits and charges still apply. Licensed under MIT.