GIDE | The coding agent that runs on your machine

Example session. The edit is waiting on you. Keep it or undo it.

02The problem

Cloud AI tools send your code away and bill you by the token.

To answer a prompt, a cloud coding tool uploads the files it needs to a provider's servers, and every token in and out is metered. GIDE runs the model on your machine instead: the code stays where it is, and local models have no per-token cost.

A cloud AI tool GIDE

your machine

auth.ts
payments.ts
schema.sql
internal-api.md

model: on the provider's serversmodel: runs here

provider's servers

tokens billed0

Illustration: the files and the counter are made up, not a measurement.

03Control

Nothing is written until you say so.

An agent that edits your repo behind your back is a liability. GIDE stops twice, once at the plan and once at every write, and records each decision on your machine.

01 · plan gate
It shows you the plan first
Which files it will change, what, and why, before any file is touched. Approve it or send it back.
02 · write gate
Then it asks before every write
Every file write and shell command stops here with the diff in front of you. A rejection goes back to the agent so it tries something else.
03 · audit log
And every decision is on disk
Plan approvals, writes and rejections land in a local JSONL audit log that stays on your machine.
Illustrative entries, written by the two gates above.

05Deployment

Your org decides where it runs.

Pick where the model runs. Everything else stays the same: the agent, the plan and write gates, folder-scoped access and the audit log on the developer's machine.

Air-gappedEnterprise

The model, your code and your prompts never leave the machine.
Enterprise: offline license activation, no sign-in and no network needed, ever.
The editor itself works with no sign-in.
Talk to us

GIDE cloudEnterprise

A GIDE-run model on GIDE's own Azure VM.
Enterprise only, and opt-in.
Ephemeral: nothing is logged or stored.
A dedicated instance is available on request.

Hybrid

Optional: a local and a cloud model stay ready together; @mention either for one message.
Cloud models run on your own key (Claude, GPT, Gemini), or use a model on your own VPS or private network. We can help set it up.
Whether cloud is allowed at all is your organisation's choice.

In every modethe same agent, plan gate, write gate, folder-scoped access and local audit log.

06Codebase

Understands your whole codebase, not just text matches.

Ask where a discount is applied. Text search hands back every line that happens to contain the word. GIDE follows how the code connects, which function calls which, and answers from that.

text search$ grep -rn discount

Every line with the word, from code, comments, CSS and the changelog alike. Nothing says which ones matter or how they connect.

applyDiscount, the three functions that call it and the one it calls.

Under the hood this is Hybrid JEPA: semantic embeddings combined with dynamic call-graph analysis, so the agent reads the parts of the repo that matter to the question.

The graph is an illustration of the idea, not a view of GIDE's internal data.

07Context

Switch models. Keep your context.

Memory, history and your conventions are local files, read by whichever model answers. Change your default, @mention a cloud model for one message, or move from the editor to the terminal: nothing is lost.

Example session editor · ornith-9b · local editor · claude · your key terminal · ornith-9b · local

editor · ornith-9b · local
cap discounts at 50%, it's a product decision
Update(src/discount.js) · asked first, accepted
memory: discounts capped at 50% (product decision)
@claude (your key) · one message, sees the conversation so far
review the discount change
The 50% cap matches the decision in memory. One gap: a negative percentage is not handled.
$ gide · same conversation, still on the local model
add a test for a negative percentage
Write(src/discount.test.js) · asked first, accepted
$ node --test · 2 passed, 0 failed

On this machine

Context Memory Servernotes, decisions and file summaries the agent keeps as it works
<project>/.gide/history/the conversation, shared by the editor and the CLI
.gide/skills/your project's conventions

Read byOrnith 9B, localClaude, your keythe gide CLI, localwhichever model answers

The Context Memory Server is structured memory outside the model: what the agent has learned about your project, kept on your machine and handed to whichever model answers.

08Your way

Works your way.

Any model, the terminal, your own tools and your own conventions. The agent and its gates stay the same.

Local by default, cloud with your key

Ornith 1.5 9B runs on your machine, and any GGUF works too. Or connect Claude, GPT, Gemini, a custom endpoint or your own VPS, with your key.

gide

$ gide models pull qwen
$ gide connect anthropic
⎿ key stored locally

Memory and context each model needs

09Regulated teams

Built for regulated teams.

The controls run on the developer's machine, and they stay the same whichever model answers.

plan gate: The agent writes a plan you approve before it touches a file.
write gate: Every change to a file asks first. Nothing is written until you say so.
tool calls: Every MCP tool call asks first, like a file write.
folder-scoped access: GIDE works only inside the folder you grant. Reaching outside it asks again, for this session only.
folder trust: GIDE asks before it works in a new folder, and the CLI keeps the same trust.
audit log: Gate decisions are recorded in a local JSONL file on the device.

Fully local by default

The model, your code and your prompts stay on the machine. A cloud model is used only when someone picks one. On Enterprise, we set it up to your requirements, including a build with no cloud options if you need one.

Air-gapped Enterprise

Offline license activation: no sign-in and no network needed, ever. Talk to us

10Questions

Questions, answered.

Can GIDE run fully offline?: Yes. The model, your code and your prompts stay on your machine. On the standard download you sign in once to turn on the AI; after that GIDE runs fully offline and only goes online when you ask it to, like a web search, or to check for updates. The editor itself works with no sign-in. On Enterprise, air-gapped deployment uses offline license activation: no sign-in and no network needed, ever. You can also connect a cloud model with your own API key (Claude, GPT, Gemini) when you want one. Read the documentation for architecture details.
Is GIDE suitable for regulated or air-gapped environments?: Air-gapped deployment, with offline license activation and no network needed, is available on Enterprise. In every setup GIDE works only inside the folder you grant, asks before it writes, and keeps a local JSONL audit log on the device. Have specific requirements? Contact us.
What does the agent ask before doing?: It writes a plan you approve before it touches a file. Every file write, shell command and MCP tool call then asks first, and reaching outside the folder you granted asks again, for this session only. Your decisions are recorded in a local JSONL audit log on the device. You can allow everything for the rest of a session if you choose.
Can we run GIDE with no cloud options at all?: Yes. GIDE runs fully local by default, and a cloud model is used only when someone picks one. On Enterprise, we set it up to your requirements, including a build with no cloud options. GIDE cloud, a GIDE-run model on GIDE's own Azure VM, is Enterprise only and opt-in.
What AI model does GIDE use?: GIDE is model-agnostic. Ornith 1.5 9B is the default local model; Qwen and Gemma are one command away, and you can point it at any GGUF file you like, all running through llama.cpp with GPU acceleration on NVIDIA (CUDA) and Apple Silicon (Metal). Claude, GPT, Gemini and any OpenAI- or Anthropic-compatible endpoint work too, with your own key. Keep a local and a cloud model ready together and @mention either one for a single message; the conversation carries over. The agentic harness is identical either way; see the memory and context matrix for what runs on your hardware.
Do I need to pay for AI tokens?: Not for local models: they run on your hardware, so there are no per-token charges no matter how much you use them. If you connect a cloud model with your own key, you pay that provider directly at their rates. New users get a 30-day free trial on any paid plan. Download and try it, no credit card required to start.