LiteLLM Local Console

LiteLLM Local Console runs a local LiteLLM Proxy from VS Code without requiring a separately installed Python, Node.js, npm, Docker, or desktop launcher.

This Windows x64 extension includes only a compact CPython core. LiteLLM and its Proxy dependencies are downloaded into VS Code per-user storage when you choose Download / Update LiteLLM. The verified Windows x64 installation used about 150 MB of downloaded/cacheable wheels and about 535 MB for the active runtime after unpacking; the retained pip cache can use a further ~150 MB to speed later updates. Dependency sizes vary by LiteLLM release. The extension never packages, imports, or uploads the developer's desktop configuration, upstream URLs, or credentials.

First run

1. Open the LiteLLM view in the Activity Bar.
2. Leave the version at 1.95.0 for the desktop-compatible release, or explicitly enter latest to install an unpatched newer LiteLLM release.
3. Select Auto probe (recommended), Official PyPI, Tencent Cloud, or Huawei Cloud.
4. Watch source probing, pip preparation, download, compatibility patching, import validation, and activation in the sidebar. The running installer can be cancelled.
5. Add an environment variable, open the complete Markdown configuration guide, then save YAML and start the local service.

The service binds to 127.0.0.1:4000 by default. It never terminates a process that already owns that port.

Guided configuration and testing

• The sidebar opens a packaged Markdown configuration guide with copyable OpenAI, OpenAI-compatible, Azure, Responses, Groq Responses-to-Chat bridge (model: groq/...; use_chat_completions_api: true is optional), Claude/Anthropic Messages request-format, Proxy-key, multi-model, and troubleshooting examples. The YAML card also offers the same examples as load-only buttons; loading never saves over the file until you click Save.
• Use model_name as the local alias, litellm_params.model as the upstream model, and api_key: os.environ/VARIABLE_NAME for credentials. Keep API keys and Proxy master_key values out of YAML and its .bak backups.
• The Environment Variables card can list and update three scopes: LiteLLM Private (VS Code Secret Storage, recommended for keys), Windows User, and Windows System. Values are masked by default, and an individual value can be explicitly shown then hidden again in the current panel. System edits require an explicit UAC administrator approval.
• The next service start reads Machine, User, and LiteLLM Private variables, so a newly saved User variable does not require restarting VS Code. Restart LiteLLM after changing YAML or variables.
• The service card tails litellm.out.log in the sidebar while it runs and can open the full file in an editor. Common key/token patterns are redacted before log text is rendered in the webview.
• The Chat Test card sends only to the currently configured local proxy at 127.0.0.1. It supports Chat Completions, OpenAI Responses API, and Claude/Anthropic Messages request format (the latter does not force a Claude upstream), plus the YAML model alias, optional system prompt, streaming output, cancellation, and an optional one-time Proxy access key that is never persisted. Transient upstream 502/503/504 responses are retried once with a visible status.
• If an upstream model needs network proxy access, the optional Upstream network proxy control securely saves HTTP_PROXY, HTTPS_PROXY, and ALL_PROXY for the LiteLLM process, preserves existing NO_PROXY entries while adding local bypasses, and restarts an extension-managed service. The local browser-to-proxy request is never routed through that proxy.
• The same control includes Clear proxy and restart service. It removes the extension-private proxy variables without starting a stopped service. The service card and Process Management card show the recorded service PID and listener PID; startup writes the PID and begins tailing the log immediately, before health checks finish.

Reliability and compatibility

• China-friendly downloads: Tencent Cloud and Huawei Cloud mirrors are probed alongside PyPI. For Chinese mirrors, only their host is added to NO_PROXY/no_proxy; the rest of the user's proxy settings remain unchanged. A failed direct mirror request retries once through the existing proxy before falling back to another source.
• Offline-friendly startup: the bundled LiteLLM model-cost map is used by default, avoiding a startup request to GitHub. Set litellmConsole.useLocalModelCostMap to false only when you deliberately want LiteLLM to fetch its remote cost map.
• Safe updates: dependencies install into a staging user directory. The extension verifies the LiteLLM version, the Proxy import, the offline tokenizer cache, and then atomically switches to the new runtime. A previous runtime is retained as a timestamped backup; an interrupted install does not overwrite it, and its owned partial staging directory is safely removed.
• Desktop compatibility: for exactly LiteLLM 1.95.0, the verified desktop web_search_preview to web_search compatibility patch is applied only when the official pre-patch file hash matches. The extension also seeds the verified cl100k_base rank table locally so tiktoken does not need an extra Azure Blob download.
• Privacy: the managed configuration starts empty and contains no API key. Use os.environ/VARIABLE_NAME in YAML for credentials. You can set litellmConsole.configPath to an explicit absolute configuration path if preferred.

Marketplace target

This package targets win32-x64. The bundled Python executable is platform-specific, so separate packages with matching Python cores are required for Windows ARM64, macOS, and Linux.

Publishing to the VS Code Marketplace requires the publisher account and a Personal Access Token owned by the publisher; neither is stored in this project.

Licenses

The extension source is MIT licensed. The bundled CPython core retains its Python Software Foundation License (included in the package); LiteLLM and Proxy dependencies are downloaded separately from the selected package index under their own licenses.