GIDE 2.3.5: the GIDE CLI, Linux builds and Folder Trust

GIDE 2.3.5 is the release where GIDE leaves the editor window. The whole agent now runs in your terminal as the GIDE CLI, sharing one local server with the app. Linux joins macOS and Windows in the same release pipeline, Folder Trust puts a check in front of every read and write, and GIDE now opens into a conversation.

At a glance

GIDE CLI. The full agent as a terminal client, sharing sessions, models, memory and the Write Gate with the editor.
Local API Server. Mint a key so any OpenAI-shaped client can use the model running on your machine.
Linux builds. The app as .deb, .rpm, .tar.gz and .AppImage, and the CLI as a Linux x64 tarball.
Folder Trust. A folder has to be trusted before the agent reads or writes inside it.
Chat-first launch. GIDE opens into a conversation with your session history.
One persistent shell. A cd or an activated virtualenv still applies to the agent’s next command.

The GIDE CLI

What changed

The whole agent now runs as a terminal client. It talks to the same local server the editor uses, so a session you start in one place is the same session in the other, with the same model, the same memory and the same Write Gate.

How to use it

# One-shot prompt
gide -p "add a retry to the fetch helper"

# Pick up where you left off
gide --continue
gide --resume

The app and the CLI at the same time

The editor and the CLI coordinate through a lease instead of fighting over the model. Starting the CLI while the app is open no longer restarts anything, and closing one doesn’t strand the other.

Everything the picker does, as commands

Models, server lifecycle, account, providers, trust and machine diagnostics are all available from the terminal:

gide models    gide server    gide login     gide connect
gide cloud     gide trust     gide hardware  gide doctor
gide usage     gide update

Installing and updating the CLI

The CLI installs with a curl installer, a Homebrew formula or a Windows installer, and gide update handles self-updates afterwards. Builds ship for macOS on Apple Silicon, Linux x64 and Windows x64.

The editor’s PATH shim was renamed to gide-app, so gide in a terminal always means the agent.

Local API Server

gide apikey create mints an sk-gide key for the local server, so any OpenAI-shaped client can point at the model already running on your machine. gide usage reports what each key spent, which on a local model is nothing.

gide apikey create
gide usage

Linux builds for the app and the CLI

Linux now ships through the same release pipeline as macOS and Windows: the app as .deb, .rpm, .tar.gz and .AppImage, and the CLI as a Linux x64 tarball. The updater reads the Linux manifests per install format, so an .AppImage is never offered a .deb update.

Folder Trust

What changed

Before the agent reads or writes inside a folder, that folder has to be trusted. Trust is stored once and shared by the editor and the CLI.

How to use it

You’ll see a dialog in the workbench or a prompt in the terminal the first time GIDE works in a folder. To manage trust by hand:

gide trust

Tighter path containment

Path containment was tightened at the same time. Sibling directories that only share a name prefix with the project root, and symlinks pointing outside it, can no longer be used to escape the project.

Chat-first launch

GIDE opens into a conversation rather than an empty editor. The new launch surface has a left rail, a hero and your session history. Session titles write themselves from the first turn, and Up and Down recall previous prompts in the chat input the way a shell does.

When a command the agent ran fails, the transcript offers Fix Error, Explain and Dismiss inline, so you can redirect the run without retyping the context.

Under the hood

One persistent shell

run_command used to start a fresh shell for every command, throwing away everything the previous one did. It now reuses a single persistent hidden shell, so a cd, an exported variable or an activated virtualenv still applies to the next command. This works in released builds, not only from source.

Model identity and native tool dialects

GIDE derives a profile from the model file itself: the prompt dialect, the tool-call format and the context budget. Tool calls are then spoken in the dialect that model was actually trained on, instead of forcing every model through one grammar. Point GIDE at any .gguf and it works the rest out.

A newer model runtime

The bundled model runtime was updated, closing a five-month gap in architecture support. Newly loadable: Gemma 4, DeepSeek V4, Mistral 4 and Cohere2-MoE, plus the Granite 4.0 and 4.1 chat templates and the DeepSeek-OCR and Hunyuan-VL variants.

Refreshed online models

The Private Cloud picker and the CLI’s gide connect now offer the current generation from each provider:

Anthropic: Claude Opus 5, Sonnet 5 and Haiku 4.5
OpenAI: GPT-5.6 Sol, Terra and Luna
Google: Gemini 3.1 Pro, 3.6 Flash and 3.5 Flash-Lite

Any other model your key can reach is still available by ID.

Windows install

Windows installer download speed is fixed, a cmd.exe entry point ships alongside the PowerShell path, and interrupted installer downloads now pick up where they stopped instead of starting again from zero.

Fixes

The welcome screen now reflects the current model immediately on connect or disconnect, and shows only one greeting.
A missing file passed to read_file no longer sends the agent into an endless retry loop.
A failed run_command halts the run and hands you the quick actions instead of continuing.
The CLI tarball no longer crashes on launch, and several Linux packaging issues were fixed.

For the short version of every change, see the GIDE changelog.