v1.5.0-rc.2

Released: July 13, 2026

The second release candidate on the 1.5 track. The headline is a unified Terminal/AI console that can launch CLI coding agents — Claude, Codex, Gemini and Copilot — directly inside Kunobi. Alongside it: three new git-provider Sources (GitHub, GitLab and Bitbucket) that answer "who has access to this repo, and why?", a substantially faster and better-looking relationship graph, and Agent Gateway — now free and built in by default — reaching all eight major LLM providers. Install from the kunobi-unstable channel to try it; the stable track is unchanged.

What's New

CLI coding agents inside Kunobi — the Terminal and AI buttons merge into a single Terminal console with two tabs. The AI tab now opens on a CLI agent picker (Claude, Codex, Gemini, Copilot) and launches the agent right in the panel, with a one-click switch to Cloud AI. Account and Settings move up into the titlebar, trimming the left strip from 7 items to 5.
GitHub, GitLab and Bitbucket Sources — three new extensions bring visibility and administration over teams, members and repository permissions into the Sources and Data pages, with a provenance graph that explains why someone has access to a repo (direct grant, team, org role, or inheritance). They also surface the security posture: branch and tag protection rules, standing credentials, 2FA enforcement, outside collaborators and pending invitations. Each provider degrades honestly where its API can't answer rather than inventing data. These are Pro-plan features.
Agent Gateway is now a free, built-in part of Kunobi — it arrives ready to use on a fresh install instead of being something you go and find in the marketplace. It stays free, and it is still yours to remove: if you uninstall it, it stays uninstalled.
Agent Gateway reaches every major LLM provider — provider support grows from 2 to 8: OpenAI, Anthropic, OpenAI-compatible, GitHub Copilot, Gemini, AWS Bedrock, GCP Vertex and Azure. That unlocks local models (Ollama, vLLM, LM Studio), the OpenAI-compatible long tail (OpenRouter, Groq, Together, Fireworks) and the three major clouds — with a custom base URL and the right auth mode per provider, including ambient machine credentials where they exist.
Virtual models with A/B routing — an Agent Gateway LLM proxy can now expose a client-facing model name that maps to one or more real provider models. One target is a plain alias (chat → gpt-4o); several weighted targets give you a per-request A/B split.
Resilient proxies — every LLM and MCP proxy can retry transient upstream failures (429, 500, 502, 503, 504) with configurable backoff, and abort requests that exceed a request timeout. Configured inline on the proxy form.
Let an AI assemble your gateway config — a connected AI assistant can now build a complete Agent Gateway setup (providers, virtual keys, MCP targets, rate limits and proxies) without ever seeing a secret. Every credential is left as a placeholder, each pending resource shows a Needs secret badge, and a proxy that depends on one refuses to start until you fill it in.
Extensions can expose tools to AI assistants — a new platform capability, opt-in per extension and off by default. Read-only tools run directly; anything that mutates requires explicit human approval first.
Table, graph, or both for Sources — credentialed sources can now be explored as a table, as a relationship graph, or side by side in a resizable split, matching the Kubernetes drilldown experience. Clicking a graph node highlights its table row and vice-versa, and ⌘⇧V switches view mode here too.
Generative dashboards are free everywhere — AI-authored dashboards and data views now work in every build, including stable and unstable. They previously required Pro in production, and the frontend gate meant no shipped binary could enable them at all.
A roomier Git card — the commit view moves the changed-file list into a resizable column and gives the diff the full height of the card, rendered side-by-side instead of stacked. A clean repo reclaims the half of the card that used to just say "no uncommitted changes", drilling into a submodule no longer pushes the dashboard below it down, and long commit messages collapse to a single line.

Improvements

A rebuilt relationship graph — the graph now renders on a new GPU canvas driven by a native layout engine, and it shows. Panning and zooming stay smooth on large graphs thanks to spatial hit-testing and level-of-detail rendering, edges are routed and labelled properly, and node cards render the same way they do everywhere else in the app.
Zoom and fit controls, on by default — the graph gets familiar zoom-in / zoom-out / fit-to-view buttons, and the mouse wheel now zooms toward the cursor.
Edges are color-coded by relation — every relation family (owned-by, manages, mounts, contains, depends-on, routes-to, …) gets its own color, and each Relations filter chip carries a matching color dot as a built-in legend. Previously every edge in a graph could collapse to the same purple.
Clearer graph selection — a single click selects a node and a double-click collapses it, so exploring no longer collapses things by accident. Marquee selection works across the canvas.
Extensions feed the graph — Agent Gateway now contributes its entities and relationships to the knowledge graph, so a connected AI assistant can traverse them like any other resource.

Bug Fixes

Table columns stop clipping their content — columns rendering monospace text, badges, status dots or icons now size to fit what is actually shown, across the Kubernetes drilldown, Cloudflare, Cloud Providers, Audit, Sources and Settings tables.
No more duplicate edges when expanding the graph — graph traversal now de-duplicates edges, so expanding a node no longer stacks repeated relationships on top of each other.
Faster graph views on large clusters — expanding a fully-visible graph no longer walks the whole descendant tree, and the card overlay stays glued to the canvas while you pan.
Assorted layout and rendering fixes — a round of repro-tested fixes across graph layout, canvas and diagram rendering, including containers that could be dissolved incorrectly and edges that landed in the wrong place.
The Upgrade button behaves like a button — it now shows a pointer cursor on hover.

Installation

This is an unstable release. To try it on Homebrew or winget, use the unstable identifiers:

brew install --cask kunobi-unstable

winget install kunobi-ninja.kunobi.Unstable

Existing installs on the stable channel are unaffected. You can also log in and download the unstable build directly, or update from within an existing unstable install.

Feedback

We'd love to hear your thoughts!