AgenticProxy for VS Code

AgenticProxy lets you use any number of OpenAI-compatible model endpoints as first-class chat models inside VS Code Chat — mixing local models (Ollama, vLLM, LM Studio, llama.cpp, LocalAI) and hosted APIs (DeepSeek, Groq, OpenRouter, Together, SiliconFlow) all in the same model picker you use for Copilot.

Works with streaming, tool/function calling, and keeps your API keys secure in the OS keychain. Built on VS Code's official Language Model Chat Provider API (VS Code 1.104+).

✨ Why AgenticProxy?

VS Code Chat natively supports only a handful of model vendors. AgenticProxy fills the gap:

• One vendor, many endpoints — unlimited OpenAI-compatible providers under a single "AgenticProxy" vendor in the model picker.
• Mix local & cloud — Ollama on one machine, DeepSeek on another, Groq on a third — all in the same workflow.
• No code, no config files — manage everything through friendly menus and a sidebar view; no JSON editing.
• Secure by design — API keys live only in VS Code's encrypted SecretStorage (macOS Keychain / Windows Credential Manager / Linux Secret Service). Keys are never written to disk or logged.

✨ Features

• Automatic Model Discovery — queries GET {baseUrl}/models for every configured endpoint, with per-provider error isolation (one offline server never breaks the others).
• Custom Model Overrides — add explicit model IDs and display names when an endpoint's /models route is missing or incomplete.
• Streaming & Tool Calling — token-by-token streaming, with tool/function calling deltas reconstructed for agentic workflows.
• Rich Sidebar View — dedicated activity-bar view: add providers, expand per-provider model lists, configure, or delete with one click.
• Connection Testing — every provider is validated against the endpoint before saving, so bad URLs/keys are caught instantly.
• Secret Rotation — update or clear API keys without re-creating a provider.

🚀 Getting Started

Installation

Add your first provider

1. Open VS Code Chat (Cmd+Alt+I / Ctrl+Alt+I).
2. Click the model picker → select Manage Language Models — or run AgenticProxy: Manage Providers from the Command Palette.
3. Select Add New Provider and fill in:
   • Nickname — a human label, e.g. Local Ollama, DeepSeek.
   • Base URL — the OpenAI-compatible root, e.g. http://localhost:11434/v1 or https://api.deepseek.com/v1.
   • API Key (optional) — leave blank for local servers without authentication.
4. AgenticProxy tests the connection automatically and registers every discovered model in the Chat model picker.

The provider form is a dedicated, persistent view that stays open when focus moves — so you can copy/paste URLs and keys across windows without losing your place.

Pick an AgenticProxy model in Chat

• Open Chat, click the model dropdown → under AgenticProxy you'll see every provider/model pair (e.g. Local Ollama: llama3.3:70b, DeepSeek: deepseek-chat).
• Switching models is instant — the prompt is routed to that endpoint at request time.

🧭 The Sidebar

AgenticProxy adds a dedicated activity-bar icon (plug ⚡). The Providers view shows:

• + Add Provider — shortcut to add a new endpoint.
• Refresh All Models — re-queries /models across every provider.
• Provider cards — nickname, base URL, model count; View Models expands the list; Configure opens per-provider management (edit / key / overrides / test / delete); Delete removes the provider and its secret key.

🛠 Commands & Settings

Commands

Settings (settings.json)

🔒 Security & Privacy

• Keys never touch disk. API keys are stored via VS Code's SecretStorage (macOS Keychain, Windows Credential Manager, Linux Secret Service) and re-hydrated only when a request is actually made.
• Deletion wipes keys. Removing a provider also deletes its key from the secret store.
• Header tampering blocked. Your Authorization header is always set from the stored key; custom headers can't override it.
• No telemetry. The extension never sends logs, key material, or prompts anywhere except the endpoints you explicitly configured.

📡 Supported Endpoints

Compatible with any OpenAI-protocol server:

• Local / self-hosted: Ollama (/v1), vLLM, LM Studio, llama.cpp server, LocalAI, text-gen webui, TabbyAPI, and more.
• Hosted: DeepSeek, Groq, OpenRouter, Together AI, Mistral, Moonshot, SiliconFlow, and any other api.*/v1 endpoint.

Some servers don't expose a complete /models list or support streaming. If yours is quirky, use Custom Model Overrides (under the provider's details menu) to explicitly register model IDs, display names, and token/tool capabilities.

⚠️ Limitations

• Requires VS Code 1.104+.
• Model lists are updated on edit/save/refresh — not auto-polled (use AgenticProxy: Refresh Models).
• Per-model capabilities (image input, tool calls) are best-effort heuristics — endpoints that don't advertise use defaults (tool calling: true, image input: false).
• max_input_tokens / max_output_tokens are heuristic; tune the settings if a model gets truncated.
• Token counting is an estimate (≈4 chars/token), not an exact tokenizer — fine for context management, not for billing.

🤝 Contributing

Issues, feature requests, and PRs are welcome — see DEVELOPMENT.md for architecture, dev commands, and publishing.

📄 License

MIT