Augment Proxy Manager

Use any AI model as the backend for the Augment VSCode extension.

Zero Injection · Zero Login · Zero Configuration

English · 中文

---

How It Works

Augment Proxy Manager runs a local HTTP proxy that intercepts the Augment extension's API requests and forwards them to your chosen AI provider.

Augment Extension  →  Local Proxy (:8765)  →  Your AI Provider API
                       ↑ Auto-configured       ↑ Viking context injection

When the proxy starts, it automatically sets augment.advanced.completionURL to point at the local proxy and augment.advanced.apiToken to a placeholder token. The Augment extension detects the config change, switches to API Token mode (bypassing OAuth), and routes all traffic through the proxy. When the proxy stops, the config is cleared and the extension returns to normal.

Supported Providers

Provider	Protocol	Default Model
Anthropic	Native	claude-sonnet-4-20250514
MiniMax	Anthropic-compatible	MiniMax-M2.2
DeepSeek	Anthropic-compatible	deepseek-chat
Google Gemini	Google Native	gemini-3-pro-preview
OpenAI	Native	gpt-4
GLM (Zhipu)	OpenAI-compatible	glm-4.7
Kimi (Moonshot)	OpenAI-compatible	kimi-k2.5
Kimi Coding Plan	Anthropic Messages	kimi-for-coding
Custom	Anthropic / OpenAI	—

Quick Start

1. Install this extension alongside the official Augment extension
2. Select a provider and enter your API key in the sidebar panel
3. Start the proxy — everything else is automatic

That's it. No code injection, no reload, no login required.

Features

v3.4 — Agent Tool System

• Tool type system — Generic Tool<Input,Output> interface with buildTool() factory, fail-closed safety defaults
• ToolRegistry — Unified tool registration, lookup (with aliases), dispatch, and concurrent partitioned execution
• Concurrency partitioning — Read-only tools run in parallel, write tools run serially
• 5 new agent tools: bash, glob, grep, file_read, list_directory
• Auto-injected schemas — New tool JSON schemas auto-injected into Anthropic/OpenAI/Gemini formats
• Unified interception — All three provider paths integrate ToolRegistry async interception

v3.0 — Intelligent Context Engine

• Viking layered context — Inspired by OpenViking: L0 summary / L1 structure / L2 full content, loaded on demand
• Directory aggregation + recursive drill-down — Vector pre-filter → directory aggregation → top directory drill-down
• Session Memory — Auto-extracts user preferences from conversations, persisted via LevelDB
• Local embedding models — 5 ONNX models (22MB–118MB), one-click download and switch from the sidebar
• Remote Embedding API — GLM / OpenAI / custom remote embedding with automatic local fallback

Proxy Core

• Zero-injection bypass — Auto-configures Augment to use the proxy without modifying any code
• Streaming responses — Real-time SSE streaming for chat, completion, and instructions
• Full Agent mode — Tool calls, file editing, codebase retrieval all work correctly
• Protocol translation — Unified Augment request format translated to Anthropic / OpenAI / Google backends
• Outbound proxy support — HTTP_PROXY/HTTPS_PROXY/NO_PROXY for corporate environments (CONNECT tunneling)
• Request correlation — Every outbound API call tagged with x-request-id for debugging
• Transient error retry — Automatic single retry on 502/503/504 and transport errors
• Hot-reload config — Switch providers or models without restarting the proxy

RAG Semantic Search

• Local code index — Built-in TF-IDF + BM25 + semantic hybrid search, no cloud sync needed
• 5 embedding models — MiniLM / BGE / E5 series including multilingual
• Per-model cache — Independent cache files per model, switching doesn't lose history

Enhancements

• OMC orchestration — Integrates oh-my-claudecode with 6 orchestration modes
• Thinking mode — Extended thinking support for DeepSeek, MiniMax, GLM, Claude, Gemini
• Prompt caching — Auto-injects cache_control for supported providers
• Context compression — Token-aware intelligent chat history compression
• Sidebar control panel — Visual UI for all configuration and runtime status

Local Embedding Models

Model	Size	Dims	Language	Notes
MiniLM-L6	22MB	384	English	Smallest and fastest
MiniLM-L12	33MB	384	English	12 layers, more accurate
BGE-Small	33MB	384	English	Good for code search
BGE-Base ⭐	109MB	768	English	Best value, recommended
E5-Multi-Small	118MB	384	Multilingual	Chinese/English/Japanese/Korean

Models use Xenova/transformers.js ONNX format, auto-downloaded on first use.

Configuration

Setting	Default	Description
augmentProxy.provider	anthropic	AI provider
augmentProxy.port	8765	Proxy server port
augmentProxy.enableContextCompression	true	Enable smart context compression
augmentProxy.compressionThreshold	80	Compression trigger threshold (%)
augmentProxy.{provider}.baseUrl	per provider	API endpoint URL
augmentProxy.{provider}.model	per provider	Model name
augmentProxy.omc.enabled	false	Enable OMC orchestration
augmentProxy.embedding.localModel	Xenova/all-MiniLM-L6-v2	Local embedding model
augmentProxy.embedding.enabled	false	Enable remote Embedding API

Provider-specific options (thinking mode, caching, etc.) are under augmentProxy.{provider}.*. API keys are securely stored in VSCode's built-in SecretStorage.

Architecture

src/
├── extension.ts              # Extension entry point
├── proxy.ts                  # HTTP proxy server + routing
├── outbound-proxy.ts         # Outbound HTTPS proxy (CONNECT tunnel) + retry
├── messages.ts               # Augment protocol parsing + system prompt injection
├── sidebar.ts                # Sidebar Webview UI
├── config.ts                 # Provider configuration
├── globals.ts                # Global state
├── context-manager.ts        # Context management
├── context-compression.ts    # Smart compression
├── omc.ts                    # OMC orchestration
├── tools/                    # v3.4.0 tool system
│   ├── Tool.ts               # Core interface + buildTool() factory
│   ├── ToolRegistry.ts       # Registry (lookup/dispatch/concurrent execution)
│   ├── StrReplaceEditorTool, SaveFileTool, ApplyPatchTool, BashTool
│   ├── GlobTool, GrepTool, FileReadTool, ListDirectoryTool
│   └── shared/               # Patch parser, path utils, input fixer
├── providers/
│   ├── anthropic.ts          # Anthropic streaming + proxy + retry
│   ├── openai.ts             # OpenAI streaming + proxy
│   └── google.ts             # Google Gemini streaming
└── rag/
    ├── index.ts              # RAG index + Viking-enhanced search
    ├── embeddings.ts          # Embedding engine (5 local + remote API)
    ├── viking-context.ts     # Viking L0/L1/L2 layered context
    ├── session-memory.ts     # Session memory (LevelDB)
    └── storage.ts            # LevelDB persistence

Cross-Platform Support

Editor	macOS / Linux	Windows
VSCode	~/.vscode/extensions	%USERPROFILE%\.vscode\extensions
VSCode Insiders	~/.vscode-insiders/extensions	%APPDATA%\Code - Insiders\extensions
Cursor	~/.cursor/extensions	%USERPROFILE%\.cursor\extensions
Windsurf	~/.windsurf/extensions	%USERPROFILE%\.windsurf\extensions

Evolution

v1.9.0  Zero-injection proxy + RAG semantic search
  ↓
v2.1.x  Kimi/GLM multi-provider + OMC orchestration
  ↓
v3.0.0  Viking layered context + Session Memory + local embedding
  ↓
v3.1.0  File editing engine rewrite + 3-provider loop architecture
  ↓
v3.3.x  OpenAI Responses protocol + Kimi tool chain + state endpoints
  ↓
v3.4.0  Agent tool system — Tool types + ToolRegistry + 5 new tools
  ↓
v3.4.1  GLM tool loop fix + outbound proxy support

License

MIT