Antigravity-Tools-LS: Agent Bridging System for Transcoding Native IDE Plugin Protocols into Standard Big Model APIs

Antigravity-Tools-LS is a program designed for Antigravity IDE deeply customized high-performance native AI protocol adaptation gateway. It breaks the conventional reverse proxy model and adopts the “pure native LS path” technology architecture, which fully takes over the native AI protocol by deeply simulating the IDE internal plug-in protocol. ls_core The complete lifecycle of a language server process provides you with a one-stop local solution for authentication injection, multi-model protocol transcoding and intelligent scheduling of multiple accounts.

The system provides standard OpenAI, Anthropic, and Gemini API This makes it easy for any third-party client that supports these standard interfaces to call them; internally, the requests are fully delegated to the native application process. This design allows the tool to hold real credentials and establish HTTPS/gRPC native connections directly with the official back-end, ensuring that all request characteristics sent are consistent with the official licensed plugin 100%, thus fundamentally circumventing the wind-control interception. The built-in automatic version alignment and industrial-grade asset synchronization engine, coupled with strict LRU instance pool management, unlock powerful Cascade Intelligence features for you while maximizing operational stability and extremely low system resource consumption in a cross-platform environment.

Function List

• Deep multi-protocol bridge adaptation: Supports the uniform and seamless conversion of native communication protocols to the standard OpenAI (/v1/chat/completionsAnthropic (/v1/messages) as well as the native Gemini API.
• 100% Native Request Characterization Simulation: Utilizes authentic identity credentials (OAuth Tokens) to direct native processes to communicate directly with the cloud with superior anti-blocking performance.
• Seamless IDE account switching with one click: Database-level bottom-layer injection technology, automatically recognizing macOS/Windows/Linux environments, accurate modification of the state.vscdbIt also supports automatic kill and restart of IDEs, completely freeing up the manual login process.
• Unlock Cascade Smart Body Mode: Native support for advanced Cascade Agent logical reasoning and complex contextual planning, as well as support for calling back referenced search and data sources in streaming output.
• Industrial-grade, fully automated asset synchronization engine: the ability to automatically scan for installers locally or pull them from the cloud (supported) .dmg、.deb、.exe、.tar.gz), manual intervention-free extraction of the underlying ls_core and security certificates.
• Multi-dimensional version of self-healing mechanism: The built-in strict alignment mechanism ensures that the underlying runtime components perfectly match the version signature features in the request header, completely eliminating 403 Forbidden errors caused by version inconsistencies.
• LRU Instance Pool Lifecycle Governance: Intelligently detects timeout processes and automatically cleans up expired or zombie LS instances through LRU elimination policies to tightly control the extremely resource-consuming overflow of memory (OOM) problem.
• Real-time system governance and status pushThe real-time data flow based on SSE (Server-Sent Events) realizes millisecond front-end display of application status, task progress, and traffic conditions. Configuration changes are written atomically, and no data is lost during power outages.

Using Help

Antigravity-Tools-LS is designed to provide you with an out-of-the-box and highly controllable local proxy service. Please read the following detailed installation and operation instructions in order to ensure that the tool runs stably and utilizes all its features.

📌 I. Rapid Installation and Deployment Processes

The system supports multiple operating platforms, and we recommend using one-click installation scripts or Docker for deployment. The system runs by default on the local 5173 Ports.

1. One-click script installation (recommended for daily use)

• Linux / macOS environment：
  Open a terminal and execute the following commands to fully automate the dependency installation and environment configuration:
  curl -fsSL https://raw.githubusercontent.com/lbjlaq/Antigravity-Tools-LS/main/install.sh | bash
• Windows environment (PowerShell)：
  Please open PowerShell as an administrator and enter and execute the following script:
  irm https://raw.githubusercontent.com/lbjlaq/Antigravity-Tools-LS/main/install.ps1 | iex

2. Docker deployment (recommended for NAS or cloud server users)
Use Docker to better isolate your environment and avoid system dependency conflicts. Execute the following command to start the service and persist the configuration to the host:

docker run -d \
--name antigravity-ls \
-p 5173:5173 \
-e RUST_LOG=info \
-v ~/.antigravity-ls-data:/root/.antigravity_tools_ls \
lbjlaq/antigravity-tools-ls:latest

3. macOS Homebrew Installation
If you're a Mac developer, you can install it directly through the Package Manager:

brew tap lbjlaq/antigravity-tools-ls https://github.com/lbjlaq/Antigravity-Tools-LS
brew install antigravity-tools-ls

📌 II. Core Functions Operation and Configuration Guide

1. Zero-allocation asset synchronization and version alignment (core)
Antigravity-Tools-LS requires the underlying ls_core The executable file and associated credentials are required for this to work. The system has a built-in “zero-configuration synchronization engine”, which by default works via the Auto (intelligent comparison) Strategy Run:

• Initiate detection: Each time it starts up, the system automatically scans the local system for an installed Antigravity IDE, and if a newer version of the locally installed version is found, the system automatically extracts its internal core components and synchronizes the version signatures to the data/ls_config.jsonTo ensure that your API requests are never officially blocked due to an outdated version.
• Hot Updates in the Cloud: If your device is a pure server (no IDE installed), the tool will call the official cloud API to automatically download the installation package for the corresponding architecture, and unzip it silently to replace it.
• ⚠️ Windows users must be aware: Since the Windows version of the installer requires a specific tool to unpack it, you must install the 7-Zip and make sure its path is added to the system PATH environment variable (recommended command) scoop install 7zip (one-click installation). Without installing 7-Zip, you can only manually extract the official package's ls_core 和 cert.pem and place it in the tools directory of the bin/ folder.

2. One-click IDE account injection and scheduling
If you need to use multiple IDE accounts for HF API calls, the system provides automated account switching capabilities.

• In the front-end Dashboard, go to “Account Management”.
• After adding your account information, the system will actively take over the underlying files state.vscdbIt supports atomic write operations to IDE configurations under macOS, Windows, and Linux. It supports atomic-level write operations to IDE configurations under macOS, Windows, and Linux.
• When switching accounts, the system will automatically perform the sequence of “force close current IDE process -> inject new account credentials -> restart IDE process automatically”. This means you no longer need to manually log out and re-scan the login code when polling multiple accounts, the whole scheduling process is highly automated.

3. API gateway agent invocation and model configuration
After successful deployment, you can point the API addresses of various third-party clients that support the OpenAI format (e.g., Chatbox, NextChat, etc.) to the local service port:

• Base URL Proxy Address：http://localhost:5173
• OpenAI Compatible Routing: Send a request to POST /v1/chat/completions, supports streaming output.
• Anthropic Compatible Routing: Send a request to POST /v1/messages。
• Currently strictly supported model IDs (request must exactly match the following string, alias mapping is not supported at this time)：
  • Gemini Series:gemini-3.1-pro-high, gemini-3.1-pro-low, gemini-3-flash-agent
  • Claude Series:claude-sonnet-4-6, claude-opus-4-6-thinking
  • Other models:gpt-oss-120b-medium
    (Note: The current version does not support image generation modeling by requesting images directly from endpoints at this time, and real-time streaming extraction of the chain of thought (Thinking) process back to the display is still in the development stage.)

📌 III. Performance Tuning and Troubleshooting Recommendations

• Avoiding Overflow of Memory (OOM) Crashes: In the case of using a program such as claude-sonnet-4-6 This advanced model performs large context or highly concurrent queries when the underlying native ls_core It can instantly allocate extremely large amounts of memory (peaks can exceed 2GB) due to the processing of complex logic.
• Mandatory for small memory devices: If your server or NAS has less than 2GB of physical memory.Must be configured with at least 5GB of virtual memory (Swap)Otherwise, the system comes with a LRU recovery mechanism. Otherwise, the system's own LRU recycling mechanism can clean up idle zombie processes, but it is too late to guard against concurrent instant memory bursts, which can lead to containers or processes directly crashing and restarting.

application scenario

1. Unified Big Model API Aggregation Calls
   By providing standardized OpenAI, Anthropic, and Gemini APIs, development teams can seamlessly integrate and tap into their own automation scripts, third-party chat clients, or on-premises toolchains with powerful modeling capabilities that would otherwise be closed off in a specific IDE.
2. Multi-account high concurrency AI service auto-scheduling
   For heavy code generation users or multi-person teams, the system can take over the credentials of multiple IDEs through database-level injection technology. Combined with the process management and switching functions, it can realize automatic polling and load balancing for multiple accounts, which perfectly breaks the bottleneck of single-account usage frequency.
3. Low-risk communication capture and cue word engineering research
   Due to the system's approach of “holding real credentials + replicating the official process request path”, the characteristics of the network requests sent are flawless. Security engineers and big model researchers can safely listen to the original communication structure and study the cue word orchestration and data distribution mechanism of the official Advanced Cascade Intelligence.

QA

1. Question: Why does it keep prompting “Core Asset Synchronization Failed” or missing executable file when running on Windows system?
   solution: This system has an underlying dependency when auto extraction is enabled on Windows. 7-Zip to parse the official installer. Be sure to check that 7-Zip is installed on your system and that the 7z.exe path has been added to the system environment variable (PATH) correctly. If you do not want to install the dependencies, please manually download the official package and unzip the ls_core 和 cert.pem file into this program's bin/ Folder.
2. Question: Why does the container crash and restart as soon as I send a dialog request after deploying with Docker on my NAS / Raspberry Pi?
   solutionThis is a typical low memory OOM (Out of Memory) phenomenon. When the tool invokes a complex model for highly concurrent replies, the underlying component may momentarily peak at more than 2GB of memory. for devices with low physical memory, you must set up at least 5GB of swap space in the host operating system to buffer the momentary memory spike.
3. Question: Does the tool support using the Thinking mode of Claude 3.7 or generating images directly?
   solution: Currently (as of v0.0.2), the ability to extract and display the Thinking process in the front-end stream in real-time has not yet been realized, and this feature has been added to the development roadmap. At the same time, the system does not support directly invoking a dedicated endpoint for image generation, but only indirectly triggering the graphic generation capability through the main model autonomously invoking auxiliary tools (Tools) during the conversation.