v1.6.0

Released: July 30, 2026

v1.6.0 promotes the 1.6 line to stable, combining the v1.6.0-rc.1 and v1.6.0-rc.2 release candidates. The theme is control and connection: choose exactly which tools an AI assistant may touch, connect Kunobi to your assistant without a terminal, and let an agent answer with a live dashboard instead of text. Plus a full GitHub Actions surface, real multi-server fleets for bare metal and network gear, and multiple accounts per source.

If you've been on the unstable channel, this update is effectively a no-op.

Heads Up — virtual GitOps kinds removed

kunobi.ninja/HelmResources and the legacy topology pipeline behind it are gone. Flux, ArgoCD and Helm drilldowns now read your cluster's real resources through the knowledge graph. The views behave the same — but if you scripted against the virtual kunobi.ninja kind names, they no longer exist.

What's New

Connect your AI client without a terminal — install, repair or remove the Kunobi MCP server for seven clients (Claude Code, Claude Desktop, Cursor, Windsurf, Gemini, Codex, Copilot CLI) from Settings → AI & MCP, with per-client status at a glance. Also available as an optional step in first-run onboarding.
No more stale MCP servers — registrations pin an exact version instead of a bare package name that npx would cache forever. Each client shows an "Update available" pill, with a section-level Update all.
Tool Access — control what AI can reach — allow-list or deny-list exactly which tools an Agent Gateway MCP proxy exposes, picking from a live-discovered checklist of each target's real tools. Denied tools vanish from the client's listing and refuse calls. Fail-safe by design: an empty policy changes nothing, and a broken one fails closed.
Connect tab — every proxy shows the real config shape and file path for Claude Code, Claude Desktop, Cursor, VS Code and Codex, each in its native format. LLM proxies now name the provider that will actually serve a request.
Publish proxies through Kunobi — one checkbox exposes a keyless MCP proxy's filtered tools through your existing Kunobi MCP install, with no per-proxy client config.
Canvas — AI answers as a dashboard — an assistant can lay any answer out as a live, multi-card view: tables, charts, stats, graphs, heatmaps, timelines, runbooks, diffs, logs and deep links into Kunobi.
Canvases are yours to keep — every canvas lands in a searchable gallery you can rearrange, rename, pin and export. Agents patch boards card by card instead of republishing, every change is snapshotted with one-click restore, and tool-backed cards refresh on their own.
GitHub Actions — three new read-only sections: Runners (your self-hosted fleet), CI (per-repo success rates, drilling runs → jobs → steps to the exact failure), and Secrets (org and repo inventory, with visibility: all flagged). Names and metadata only — GitHub never returns secret values.
Redfish and SNMP become fleets — deep-monitor any number of servers or devices at once. Entity views (Fans, Temps, Drives, Interfaces) aggregate across the whole set with a Server or Device column. Redfish identity moved off unreliable hardware serials onto host:port, so distinct BMCs stop merging.
Multiple accounts per source — NATS, GitHub, GitLab, Bitbucket, Ceph and Cloudflare can each connect several times at once, one per tab, like Kubernetes clusters. Connecting now auto-selects and loads data — no separate step.
NATS monitoring — eight drillable kinds (Server, Connections, Accounts, Streams, Consumers, Routes, LeafNodes, Gateways) with a topology graph, plus a Latest messages browser on any JetStream stream. This is a Pro-plan feature.
Classic Helm in the knowledge graph — helm install releases now declare their resources like Flux and Argo already did, repointing on upgrade or rollback.

Improvements

Approval popups explain themselves — pending AI writes show structured parameters, the tool's description, and whether it's a dry run. "View Details" opens the actual request.
A usable audit trail — 5,000 records instead of 50, reachable by filter, search and export; partial views say so. Export writes the complete log as JSON or CSV through a native Save dialog — it previously produced no file and no error.
Temp auto-approval works — now a real button, and the configured default duration is finally honored. Numeric settings enforce their minimums, closing a case where 0 wiped the audit log.
Contextual tips for the write firewall — explaining that AI reads by default and risky actions wait for approval.
Entity detail sidebar for sources — matching the Kubernetes drilldown, and no longer resetting its tab when polling refreshes rows.
Git-provider access explains itself — each repository shows why someone can reach it: direct grant, team, or org role.
Cross-cluster ArgoCD relationships — Applications managing resources on another cluster now relate correctly.
Safer extension updates — transaction-safe installs preserve your config, and a stable code-signing identifier stops macOS re-prompting for keychain access.
Unified footer popovers, log-viewer copy, version-aware terms acceptance, and an in-app About page.

Bug Fixes

Extension updates no longer crash the app — a hook-slot collision could unmount the whole window, taking the crash reporter with it. Widgets now degrade to an inline chip, and "Update all" runs sequentially.
Extension Data pages no longer open blank — Ceph, GitLab, Bitbucket, GitHub and Redfish render their tables on the default view, with a loading spinner instead of a premature "No resources", and a proper connect prompt when nothing is connected.
Kubernetes columns behave — visibility auto-seeds only on real overflow, stored choices survive, and the namespace filter survives a kind change.
Sources shortcuts run the right action — the keys were firing the wrong actions; Escape clears selection again.
The connection picker opens where you are — no more landing on the Kubernetes tab in a GitHub or NATS workspace, and trailing sources are reachable again.
Agent Gateway edit safety — dangling references are rejected on save, optional numerics can be cleared, and editing a target no longer drops secrets you didn't mention.
A graph data-loss bug — the background sweep could overwrite a real entity with a placeholder.
Source fixes across Redfish, Ceph, SNMP and Cloudflare — cached clients, immutable names, status dots, DNS priority and surfaced errors.
Audit honesty — Clear reports the true total instead of the filtered count, and dry runs read as Success.
Polish — graph zoom, brand cards, unfrozen embedded pages, query-string deep links, and a readable legal notice on the home screen.

Installation

Feedback

We'd love to hear your thoughts!