One AI code editor for local and cloud models

If you want AI help in your editor today, you usually pick a side. One side is the AI code editor for cloud models: a polished agent, but every prompt and every file it reads goes to someone else’s server, and the bill grows with every token. The other side is a local model runner or a bare local LLM IDE plugin: your code stays on your machine, but you get a chat box and a model, not an agent you would trust to edit a real project.

GIDE is built so you don’t have to pick. It’s a hybrid AI code editor: one AI code editor for local and cloud models, and for models on your own server. Every model works inside the same agent harness, behind the same approval gates. This post explains why that harness matters more than any single model, and where GIDE is and isn’t the right fit.

The choice most developers are stuck with

The cloud-only route has two costs. The first is money. Online models are billed by the tokens you send and receive, and an agent sends a lot of them: it reads files, runs tests, reads the output, and tries again. The second is control. Your source code, your prompts and whatever the agent decides to read all leave your machine. That’s fine for a side project, and often not fine for a client codebase.

The local route fixes both, and then stops. Running a model on your own hardware means no per-token bill and nothing leaving the machine. But a model is not an agent. Without something around it that plans, edits files, runs commands and checks its own work, you are back to copying code out of a chat window.

What’s missing in both cases is the same thing: a harness you can trust, that doesn’t care where the model runs.

Why the harness matters more than the model

“Agent harness” has become the usual name for everything around a model that turns it into a coding agent: the tools, the loop, the memory and the rules about what the agent may do. Models change every few months. The harness is what you actually live with every day, and it decides whether you can hand the agent a real task.

In GIDE, the harness belongs to the workspace, not to the model. Whichever model answers, it works behind the same controls.

Plan Gate

Before the agent touches a file, it writes a plan and you approve it. A plan from a model on your laptop and a plan from an online model go through the same gate, so you always see what’s about to happen before it happens.

Write Gate

Every file write pauses for you to Accept, Reject, or Allow All for the rest of the session. A write you don’t answer is rejected after 2 minutes, and the agent is told when a write is rejected so it doesn’t keep retrying the same edit. The GIDE 2.0.0 release notes cover how it behaves in detail.

Command Gate

Shell commands wait for your explicit approval before they run, and so does every MCP tool call. You see the exact command and the working directory inline in the chat. If the session disconnects, any pending approval is rejected rather than left hanging.

Folder Trust and Path Gate

A folder has to be trusted before the agent can read or write inside it, and that trust is shared by the editor and the CLI. Path containment stops symlinks and sibling folders with a similar name from being used to escape the project. When a tool does need something outside your trusted folder, the Path Gate asks you instead of quietly refusing or quietly allowing it, and reads and writes are separate consents. GIDE 2.3.5 introduced Folder Trust and GIDE 2.3.6 added the Path Gate.

Memory that outlives the model

Memory, history and project conventions are local files that whichever model answers can read. When you switch models, the context carries over, so the next model doesn’t start from zero. On the Business plan, the Context Memory Server adds structured memory that persists across sessions: architecture decisions, naming conventions and past discussions.

A local audit log

Plan approvals, writes and rejections are recorded in a local JSONL audit log that stays on your machine. It’s on every plan. It’s how you show what the agent did, to a reviewer or to yourself.

On top of the gates, the agent checks its own work with an edit judge, post-edit verification and a dependency guard. None of this changes when you change models, and that’s the point.

Local, online and your own server in one editor

The harness is only half of it. The other half is where the model runs, and in GIDE that’s your call, per message if you want.

Local models

GIDE ships with a default local model, Ornith 1.5 9B, and local inference is unlimited on every plan. The Model Hub searches Hugging Face or a list of recommended models, picks the largest version that fits your memory, checks the download and tests the model before marking it ready. You can also point GIDE at any .gguf file you already have.

Online models on your own key

Bring your own API key for Anthropic, OpenAI or Google Gemini. Keys are stored locally in your system keychain and only sent to the provider you choose. GIDE ships no key of its own and never proxies your prompts. Custom Endpoints connect anything that speaks the OpenAI or Anthropic wire format, such as a vLLM or Ollama server or a corporate gateway.

A model on your own server

With VPS & Private Servers, you run a self-hosted model on a VPS, VPN or tailnet you control and connect to it like any other model, with your prompts staying on that network. If you don’t have a model server yet, Your Server → Set up a server writes the steps for your machine, and one command on the server installs and starts it. The VPS guide has the full walkthrough.

Using them together

This is where having all three in one editor pays off. When memory allows, two local models stay loaded and switching between them is instant. Type @ to send one message to any connected local or online model without changing your default, and the model you address sees the whole conversation.

That makes a practical workflow possible: plan the hard turn with a frontier model on your own key, approve the plan at the Plan Gate, then let a local model carry out the long tail of edits and test runs. You pay per token only for the turns where the frontier model makes a difference. We walk through it step by step in Multi-model in GIDE is more than switching models.

Who it’s for

Individual developers

If you want an agent that can actually edit your project, but you don’t want your code on someone else’s server or a token bill that tracks how hard you worked this month, GIDE gives you a local default with no per-token charges, and after a one-time sign-in it works offline. When a task needs a frontier model, it’s an @ away on your own key, and you see exactly what it costs you.

Teams in regulated industries

If you need a private AI coding assistant because your code can’t leave your network, run local models or a model on your own server and keep prompts inside your own environment. In local mode your code never leaves your machine, and GIDE never routes it through its own servers. The local audit log records what the agent did, and on Enterprise there is air-gapped deployment with offline license activation, plus SSO/SAML and custom compliance reporting.

To be clear about what that means: GIDE gives you control over where code and prompts go, and a record of what the agent did. Whether a setup meets a specific standard, such as HIPAA or defense requirements, is a decision for your own compliance team. GIDE doesn’t claim a certification on your behalf.

When GIDE might not be right for you

Your hardware is below the floor for local models. Local models need an Apple Silicon Mac with an M2 or newer and 16GB of unified memory, or a Windows PC with an NVIDIA RTX 3060 (12GB) or better. Below that, you can still use online models on your own key, but you lose the main reason to run locally. Check the system requirements first.
You expect a 9B local model to match a frontier model. It won’t on the hardest problems. The model guide puts the floor for reliable multi-step editing at 9B, and Ornith 1.5 9B sits at that floor. That’s why the plan-online, edit-locally workflow exists.
You want an agent that runs unattended. The gates are deliberate. Every plan, write and command waits for you. Allow All relaxes the Write Gate for a session, but if you want an agent that works for an hour with nobody watching, GIDE is built the other way.
You need air-gapped from day one on the standard download. The standard download asks you to sign in once to turn on the AI. After that it runs offline and only goes online when you ask it to, for example for a web search or an update check. Fully air-gapped deployment with no sign-in is an Enterprise feature.
You’re happy with a cloud-only editor. If cost and code location genuinely don’t matter for your work, the main problems GIDE solves aren’t your problems.

FAQ

Can I use local models and Claude in the same editor?

Yes. Connect Claude on your own Anthropic key and keep a local model loaded at the same time. Picking a local model doesn’t disconnect your online ones, and @ sends a single message to either one with the full conversation behind it.

Does a local AI code editor send my code to the cloud?

Not in GIDE’s local mode. The model runs on your machine and your code, prompts and responses stay there. GIDE does collect usage telemetry, such as token counts by mode, but no code, prompts, file names, file paths or AI responses. The details are on the trust page.

What is an agent harness in AI coding?

It’s everything around the model that makes it a coding agent: the tools it can call, the loop that plans and checks its work, its memory, and the rules about what it’s allowed to do. In GIDE the harness is the same for every model, so the Plan Gate, Write Gate, Command Gate, Folder Trust and audit log behave identically whether the model is local, online or on your own server.

How do approval gates work in an AI coding agent?

The agent stops at each risky step and waits for you. In GIDE the Plan Gate shows the plan before any change, the Write Gate pauses every file write for Accept, Reject or Allow All, and the Command Gate shows each shell command and MCP tool call before it runs.

Can I run a coding model on my own server and connect GIDE to it?

Yes, with VPS & Private Servers. Point GIDE at a model on a VPS, VPN or tailnet you control, or use Your Server → Set up a server if you don’t have one yet. Any server that speaks the OpenAI or Anthropic wire format also works as a Custom Endpoint.

When should I use a local model versus a cloud model for coding?

Use the local model for the long tail: exploring code, small edits, test runs and fixes, where per-token billing adds up and the work is well defined. Bring in a frontier model on your own key for the hard turns, such as planning a tricky refactor or reviewing the result.

What hardware do I need to run a coding model locally?

An Apple Silicon Mac with an M2 or newer, 16GB of unified memory and macOS 13 or later, or a Windows 11 PC with an NVIDIA RTX 3060 (12GB) or better. On Linux, GIDE ships as .deb, .rpm, .tar.gz and .AppImage.

Can GIDE work offline?

Yes, after a one-time sign-in. On the standard download you sign in once to turn on the AI, and after that GIDE runs offline and only goes online when you ask it to, for example for a web search or an update check. The editor itself works with no sign-in. Fully air-gapped activation, with no sign-in and no network at all, is an Enterprise feature.

Try it

If you’ve been choosing between an agent you can’t fully trust with your code and a model you can’t hand a real task, try running both in one place. GIDE runs on macOS, Windows and Linux, and every plan (Developer, Business and Enterprise) starts with a 30-day free trial, no credit card needed.

Download GIDE, or start with What is GIDE? for the short tour.