# GenerativeIDE (GIDE)

> A local-first agentic AI code editor built on VS Code that runs LLMs entirely on your machine with no token costs, a write gate on every edit, and support for any GGUF model.

GenerativeIDE (GIDE) is a local-first agentic code editor built on VS Code OSS that runs large language models entirely on your machine via llama.cpp, with no per-token charges and no code leaving your device. It ships with Ornith 9B bundled so the editor is productive on first launch, and supports Qwen, Gemma, DeepSeek, and any GGUF file alongside cloud models (Claude, GPT, Gemini) via bring-your-own-key. The current release is v2.3.6, available for Windows, macOS (Apple Silicon), and Linux.

## What It Is

GIDE is an agentic IDE in the same category as Cursor or GitHub Copilot, but designed around a local-first, offline-capable model. The agent accepts a plain-English goal, plans the work, edits across multiple files, runs shell commands, and verifies its own output — all without single-line completions. A "write gate" pauses every file write and shell command for user approval before anything is applied, and a rejection is fed back to the agent so it can try a different approach rather than silently failing. Every tool call, write, and approval is recorded in a local JSONL audit log.

## How the Agent Works

The agent operates in a multi-step loop:

- **Plan → Edit files → Run commands → Verify** — the full cycle runs end to end on a stated goal
- **Semantic retrieval (RAG)** — Hybrid JEPA embeddings index the whole repo so answers are grounded in real files, not guesses from file names
- **Persistent project memory** — context, architecture decisions, and naming conventions survive across restarts per project
- **Skill files** — drop a markdown file in `.gide/skills` and the agent loads it deterministically by `@mention` or file glob
- **One persistent shell** — a `cd`, exported variable, or activated virtualenv carries forward to the next command in the same session
- **Universal Refactoring Engine** — AST-level rename, extract, and restructure via Tree-sitter across 21 languages, applied atomically across files

## Model Access and Local Inference

GIDE is model-agnostic. The bundled Ornith 9B requires no setup; Qwen, Gemma, and DeepSeek are one command away; and any `.gguf` file is auto-detected — GIDE derives the prompt dialect and tool-call format from the model file itself. Inference runs through llama.cpp with GPU acceleration on Apple Silicon (Metal) and NVIDIA (CUDA). The site states Ornith 9B needs 12 GB VRAM on Windows or 18 GB unified memory on Mac. Cloud models (Claude, GPT, Gemini) work with a user-supplied API key stored in the OS keychain; GIDE never proxies prompts or marks up tokens.

## Deployment Model and Interfaces

GIDE offers three interfaces that share one local server:

- **Editor** — a full VS Code–based IDE with a chat panel; compatible with Open VSX extensions and existing VS Code keybindings
- **CLI** — the same agent in the terminal, with real exit codes so it pipes and composes like any shell tool; supports `gide --continue` and `gide --resume`
- **Local OpenAI-compatible API** — an endpoint bound to `127.0.0.1:41337/v1`; mint an `sk-gide-` key and point any OpenAI SDK at it; traffic never leaves the machine

The CLI and editor coordinate through a lease and can run simultaneously, sharing the same model, memory, and write gate.

## Target Audience and Regulated Environments

The contact page states GIDE supports developers in defense, healthcare, financial services, and robotics — environments where code privacy and air-gap operation are non-negotiable. Enterprise features include air-gapped deployment, SSO/SAML (Okta, Azure AD), custom compliance reporting, JSONL audit telemetry, bearer-token authentication, folder trust (path containment blocks symlink escapes), and 20 safety guards that prevent the agent from writing outside the project root or deleting untracked files.

## Update: v2.3.6

Version 2.3.6 is the current release. Notable additions in this version include: a one persistent shell (so environment state carries across commands), folder trust (stored once, shared by editor and CLI), support for any GGUF with auto-detected model identity, a current llama.cpp runtime tracking upstream closely (Gemma 4, DeepSeek V4, Mistral 4, and Cohere2-MoE among recent additions), and the full Gide CLI as a terminal client sharing the same local server as the editor.

## Features
- Agentic coding (plan → edit → run → verify)
- Write gate on every file write and shell command
- Local JSONL audit log
- Bundled Ornith 9B model (no setup required)
- Support for Qwen, Gemma, DeepSeek, and any GGUF model
- BYOK for Claude, GPT, and Gemini
- Semantic retrieval (RAG) across whole codebase
- Persistent project memory across sessions
- Skill files for team conventions
- Universal Refactoring Engine (AST-level, 21 languages via Tree-sitter)
- One persistent shell (environment state carries across commands)
- Folder trust with path containment
- Full MCP client with gallery and management UI
- Context Memory Server
- GIDE Wiki (architecture graph via Tree-sitter import analysis)
- Local OpenAI-compatible API endpoint (127.0.0.1)
- CLI with real exit codes and scriptable interface
- GPU acceleration (Metal on Apple Silicon, CUDA on NVIDIA)
- Air-gapped deployment support
- SSO/SAML (Okta, Azure AD)
- 20 safety guards
- Open VSX extension compatibility
- Sandboxed shell

## Integrations
llama.cpp, VS Code OSS, Open VSX marketplace, Claude (Anthropic), GPT (OpenAI), Gemini (Google), MCP (Model Context Protocol), Okta, Azure AD, OpenAI-compatible APIs, Tree-sitter, CUDA, Metal (Apple Silicon), Vulkan

## Platforms
WINDOWS, MACOS, LINUX, ANDROID, API, VSC_EXTENSION, CLI

## Pricing
Subscription-based

## Version
v2.3.6

## Links
- Website: https://generativeide.com
- Documentation: https://generativeide.com/docs
- EveryDev.ai: https://www.everydev.ai/tools/generativeide-gide