Sovereign AI
Xinity Now Ships an SBOM With Every Release
A new vulnerability is published on a Friday afternoon. Someone in your security team asks the question every regulated organisation knows: are we affected?
From this release on, the answer for Xinity is a document you already have, not an email thread with your vendor. Every Xinity release now ships with a software bill of materials. We have also cut the number of external packages the platform depends on, and added setup pages for five more AI coding agents that run against your own on-premise AI infrastructure.
An SBOM with every release
A software bill of materials (SBOM) is a complete, machine-readable list of the components a piece of software is built from: every library, every package, every version. Each Xinity release now carries one, and every release from here on will too. It uses CycloneDX, the open SBOM standard maintained under OWASP, so the scanners and dependency tools your security team already uses can read it directly.
For the people who sign off on infrastructure, this changes three moments:
When a new vulnerability is published. Your team checks the SBOM against it directly, instead of waiting for us to answer a ticket.
During procurement. The component list a security questionnaire asks for is already attached to the release you are evaluating.
During an audit. Each SBOM belongs to one exact release, so you can show which components were running at any point in time.
SBOMs are also becoming a regulatory expectation in Europe: the EU Cyber Resilience Act requires manufacturers of products with digital elements to draw one up. For Xinity, it is now part of how every release is built, not a document we produce on request.
Fewer external packages
An SBOM is only as reassuring as the list inside it. External packages are code written and maintained by third parties, and each one is something you have to trust and monitor. This release brings an extensive reduction in the external packages Xinity depends on: less to trust, and less to check when the next advisory comes out.
The same release moves the dashboard to a new version of SvelteKit, the web framework it is built on, patches the published security advisories in better-auth, the authentication library behind the login, and replaces MJML, the library we used for email templates.
Setup pages for five more coding agents
Coding agents are AI assistants that write, change and explain code inside a developer's workflow. Developers want them. Security teams want prompts and source code to stay on their own servers. Connected to Xinity, a coding agent uses the models your organisation runs on its own infrastructure.
The Coding Agents section of the Xinity docs now has setup pages for Pi, Cline, Crush, Goose and Qwen Code, alongside the existing OpenCode page. Each page takes a developer to a working agent in three steps:
Pick the models. Choose which of your organisation's chat models with tool support the agent can use, and which one it starts with.
Create the key. Create an API key for the agent right on the page. It goes through the normal key creation and permission check, so no one gets a shortcut around your access rules.
Paste the config. The page generates the configuration file the agent reads, with your gateway URL, the model names and their context windows already filled in.
The context window is how much text a model can process at once. The page takes it from the model data, and during a canary rollout (a staged update where two model versions run side by side) it uses the smaller of the two, so an agent never plans for more context than every deployment behind the model can serve. Each page also names the agent version it was last tested with, and the OpenCode page now uses OpenCode's current config format, with output limits per model.
Also in this release
Gateway (the entry point that routes requests to your models): a background response now holds its reserved backend capacity until generation ends.
Daemon (the service that runs on each server): a model download that stalls before it starts reading now ends as stalled, instead of waiting.
Dashboard: the package image keeps its example.env configuration file when it is built.
Dependencies: the better-auth advisories are patched and MJML is replaced, as described above.
Get it
Xinity is open source. The release, its SBOM and the full changelog are on GitHub. To install the CLI:
On Linux or macOS:
curl -fsSL "https://github.com/xinity-ai/xinity-ai/releases/download/v0.26.5/install.sh" \ | bash
curl -fsSL "https://github.com/xinity-ai/xinity-ai/releases/download/v0.26.5/install.sh" \ | bash
curl -fsSL "https://github.com/xinity-ai/xinity-ai/releases/download/v0.26.5/install.sh" \ | bash
On Windows (PowerShell):
irm "https://github.com/xinity-ai/xinity-ai/releases/download/v0.26.5/install.ps1" | iex
irm "https://github.com/xinity-ai/xinity-ai/releases/download/v0.26.5/install.ps1" | iex
irm "https://github.com/xinity-ai/xinity-ai/releases/download/v0.26.5/install.ps1" | iex
Running AI in a regulated organisation and want to see how this fits your setup? Book a call with our team.
AI declaration
This article was drafted with the help of AI and reviewed by the Xinity team before publication. The cover image was created with AI assistance.