GIDE 2.3.6: Custom Endpoints, Ornith 1.5 and Path Gate

GIDE 2.3.6 widens the set of models you can use and tightens how the agent asks for access. Custom Endpoints let you bring almost any model server with your own key. Ornith 1.5 9B becomes the default local model. And when a tool reaches outside your project, Path Gate now asks you instead of refusing.

At a glance

Custom Endpoints. Connect any endpoint that speaks the OpenAI or Anthropic wire format, in the app and the CLI.
Ornith 1.5 9B. The new default local model, with a longer training context.
Ask GIDE about GIDE. The assistant can look up GIDE’s own documentation when you ask a product question.
Path Gate. Access outside your trusted folder becomes a prompt, with separate consent for reads and writes.
First-run tour. A new install opens into a guided tour.
Settings for Imported Models. Temperature and GPU-layer settings now apply to any .gguf you added yourself.

Custom Endpoints

What changed

GIDE no longer only talks to the providers it ships presets for. You can connect any endpoint that speaks the OpenAI or Anthropic wire format with your own key, stored locally. That includes a self-hosted vLLM or Ollama server, a corporate gateway, a regional deployment, or a provider that didn’t exist when this release was built.

How to use it

Point GIDE at the base URL, give it a key, and pick a model from the list the endpoint reports. Saved endpoints can be added, switched between and forgotten one at a time, and the same ones appear in the app’s model picker and in the CLI:

gide connect

GIDE still ships no key of its own and never proxies your prompts.

Ornith 1.5 9B is the new default

The bundled default local model moves from Ornith 1.0 9B to Ornith 1.5 9B, and the shipped quantization from Q4_K_M to Q5_K_M (6.64 GB).

Benchmark	Ornith 1.0 9B	Ornith 1.5 9B
SWE-bench Verified	69.4	70.6
SWE-bench Pro	42.9	47.5
Terminal-Bench 2.1	43.1	46.2
Training context	32K	256K

Existing installs are offered an update rather than a fresh download, and the model is verified by SHA-512 before it’s put in place. Ornith 1.0 is no longer offered.

Ask GIDE about GIDE

Questions about the product itself now have an answer. The assistant can retrieve GIDE’s own documentation, covering chat and shortcuts, the CLI, memory, model settings, local and online models, the Plan Gate, skills, the Write Gate, and troubleshooting.

This is retrieval on demand, not a document dump. Nothing is added to the system prompt, so if you never ask a product question you pay nothing for the feature. The corpus is GIDE’s shipped docs and is fixed. It never includes your project’s files.

Path Gate: access outside your project

What changed

When a tool reached for a path outside your trusted project folder, GIDE used to refuse and hand the refusal to the model, which usually just tried again. That escape is now a prompt: GIDE classifies it and asks you.

How it works

A grant lasts for the session only. It’s held in memory and never written to your trust store, because that store matches by ancestry, and recording a one-off escape there would quietly widen Folder Trust for every future session.
Reading and writing are separate consents, so approving a read outside the root doesn’t approve a write to it.
A path you reject isn’t asked about again in the same turn.
Genuinely sensitive targets stay refused.

The gate shows in the workbench chat and prompts in the CLI.

First-run tour

A new install opens into a guided tour in the chat welcome instead of an empty surface. It covers what GIDE needs, whether your machine can run it, and how to get a model installed, with live progress and confirmation ticks as each step completes.

Smaller improvements

Settings for Imported Models

Temperature and GPU-layer settings now apply to Imported Models, meaning any .gguf you pointed GIDE at yourself rather than one of the shipped families. Changing a setting shows Applying, reloading model and then Applied in the same status-bar slot. Reset to recommended is disabled when nothing differs from the recommendation.

A reload is guarded against in-flight turns and overlapping reloads, so changing two settings quickly can’t leave the model half-loaded. The CLI applies these settings without tearing down the terminal UI.

Rail and history

Fifty chat sessions are kept instead of twenty-five. Rail scrolling and layout were fixed, the chat list updates live as sessions change, and opening an editor exits chat mode rather than leaving both competing for the same surface.

Trust prompts

The CLI trust prompt is now arrow-key driven and reflows live when you resize the terminal. The IDE dialog names the folder it’s asking about, and a refusal tells you what to do next. The trust banner stays visible in chat-first mode, and GIDE warns you when folder-trust sync to the CLI fails instead of failing silently.

Release notes on Discord

Release notes are now posted to Discord automatically when a release goes out.

Fixes

Clean installs couldn’t download a model. The downloader didn’t follow a relative redirect, which Hugging Face now returns, and the model manifest pointed at a renamed Hugging Face organization. Both are fixed.
Windows login after a reinstall. Logging in after a Windows reinstall failed with EPERM. The session file is now written to a temporary file and renamed into place.
Online models and Custom Endpoints. Saved endpoints disappeared when you deleted one or switched to a local model, and the online provider list offered local-server presets that didn’t belong in it. Key validation, live model lists, error messages, reconnect, popup dismissal and the stale-key prompt were all corrected. A provider error is now shown as the provider wrote it, and an unreachable host says the host is unreachable instead of blaming your API key.
CLI and Windows polish. /clear in the CLI now hard-clears the window and releases the server session. A console window no longer flashes on Windows during a GPU-layer reload.

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