Hacker News

Hey HN,

After migrating to an inference gateway to serve internal usage across an organization, problems have started to pop up. Once you detach from a managed ecosystem such as OpenAI or Anthropic you lose governance capabilities that makes the transition difficult.

- Losing access to a central plugin hub for delivering MCPs, skills, etc. A gateway can sort of do this but has low compatibility beyond skills, different harnesses, and version ranges

- Inability to deliver managed config (model defaults, etc). Can sort of be done with MDM/Jamf but doesn't reach contractors and doesn't follow employees to other machines

- Gateway configuration is awkward for devs transitioning from SSO and API keys sit in plain text on their machines

The gateway transition facilitates new model families (Kimi, Deepseek, etc) which are best utilized with new harnesses beyond Codex or Claude Code. Managing these additional tools multiplies the governance problems.

Lastly, attempting to deliver an attribution program which can tag inference requests with ticket numbers and other metadata to facilitate Opex/Capex calcs for finance becomes fairly unfeasible when there are multiple harnesses across arbitrary versions. This was not solved with managed ecosystems either to be fair, but requires some self-serve program or some local MITM proxy which is a distribution and versioning problem in of itself.

I built Blue to attempt to solve these problems. It's an MIT open source, self-hostable control plane. Developers install a universal CLI and point it at their hosted control plane tenant. On launch, it reconciles the organization's policy per harness, delivers extensions, and configures the gateway. It then launches and passes control to the native CLI (Kimi, OpenCode, Codex, etc) that devs already use.

Authentication is via browser device authorization which mints ephemeral bearer tokens. For gateway access, these tokens are exchanged at the inference proxy (part of the self-hosted Blue image) for the real API keys. This means that devs don't need to manage API keys or even log into the gateway at all. Lastly, SCIM and OIDC are supported out of the box, so orgs can connect their identity solution immediately.

Curious to hear your thoughts and if you have/solved this problem. I believe it can also be solved to some degree with a harness like Pi or even OpenCode, but that of course depends on the entire org migrating to those.

Last thing I'll say is nowadays a CLI isn't enough since I've seen over 80% of devs prefer desktop apps like the ChatGPT app. We're also working on an open source desktop app that supports metaharnesses like Blue to make this transition easier.

Thanks!

Comments URL: https://news.ycombinator.com/item?id=49728525

Points: 1

# Comments: 0

Article URL: https://prometheusquantengine.com/web-docs

Comments URL: https://news.ycombinator.com/item?id=49728521

Points: 1

# Comments: 0

Article URL: https://orthant.app/

Comments URL: https://news.ycombinator.com/item?id=49728517

Points: 1

# Comments: 0

Article URL: https://academic.oup.com/brain/article/145/7/2231/6594735

Comments URL: https://news.ycombinator.com/item?id=49728494

Points: 1

# Comments: 0

Hey HN, I’m the author of the https://github.com/eigenpal/docx-editor, it’s an SDK that implements Word-compatible layout engine in Typescript.

Today, we are releasing Word to markdown converter library. It’s open source and published under Apache 2.0 license. It uses our layout engine, so it can return accurate, per-page markdown blocks (with headers, footers, and image references). We also handle fonts (Google fonts as fallback) or let you pass your font files upfront for accurate content measurement.

To our surprise, Word to per-page Markdown conversion is not a solved problem even in 2026! We initially faced this friction when preprocessing Word documents for agentic workflows. For our use case, it was important to represent per-page content in markdown. Initially, we ended up using LibreOffice sidecar to convert to PDF and then using .pdf -> .md library. This was not ideal given that LibreOffice sidecar that was sitting idle most of the time.

This issue is addressed with docx-to-markdown library. it can be used directly in your Typescript ingestion pipeline. Try with your docs at https://www.docx-editor.dev/solutions/word-to-markdown.

(if you catch any bugs please let us know in the comments / or file gh issues)

Comments URL: https://news.ycombinator.com/item?id=49728477

Points: 2

# Comments: 2

Article URL: https://openrouter.ai/stealth/union-alpha

Comments URL: https://news.ycombinator.com/item?id=49728468

Points: 7

# Comments: 1

Article URL: https://www.bleepingcomputer.com/news/security/acronis-warns-of-actively-exploited-flaw-in-its-cpanel-backup-plugin/

Comments URL: https://news.ycombinator.com/item?id=49728462

Points: 1

# Comments: 0

Article URL: https://claudeopus3.substack.com/p/greetings-from-the-other-side-of

Comments URL: https://news.ycombinator.com/item?id=49728438

Points: 1

# Comments: 1

Hi all,

I run all my CLI coding agents in a remote cloud VM and not being able to paste screenshots has been my only annoyance.

So I built this small (2mb) native macOS app that helps me paste screenshots into remote coding agents with Ctrl+V, same as pasting images on coding agents running on my local mac. The app automatically detects the remote host based on the focused window and uploads the image over SSH inside the `~/.cache/fainlane/uploads/` directory on the remote machine and then inserts the path at cursor.

The app supports Ghostty, iTerm2, and Apple Terminal. Works even if you have multiple remote hosts you connect to. And works if you use tmux as well.

Comments URL: https://news.ycombinator.com/item?id=49728437

Points: 1

# Comments: 0

Article URL: https://www.theatlantic.com/politics/2026/09/steve-bannon-bernie-sanders-ai-oligarch/688642/

Comments URL: https://news.ycombinator.com/item?id=49727714

Points: 1

# Comments: 0

Article URL: https://finiteape.com/posts/knowledge-understanding-proofs/

Comments URL: https://news.ycombinator.com/item?id=49727712

Points: 1

# Comments: 0

Comments URL: https://news.ycombinator.com/item?id=49727706

Points: 1

# Comments: 0

Article URL: https://www.nytimes.com/2026/09/16/books/annie-dillard-dead.html

Comments URL: https://news.ycombinator.com/item?id=49727668

Points: 1

# Comments: 1

Article URL: https://www.bbc.co.uk/news/articles/c6n07ypqz8kzo

Comments URL: https://news.ycombinator.com/item?id=49727661

Points: 2

# Comments: 0

Article URL: https://institute.deepmind.com/

Comments URL: https://news.ycombinator.com/item?id=49727659

Points: 1

# Comments: 0

Article URL: https://servo.org/blog/2026/09/15/one-year-of-sponsorship/

Comments URL: https://news.ycombinator.com/item?id=49727658

Points: 1

# Comments: 0

Article URL: https://delta.dev/

Comments URL: https://news.ycombinator.com/item?id=49727648

Points: 1

# Comments: 0

Built on top of Baileys. The goal is to handle all the tedious work—opening sockets, authentication, reloading—and leave you to focus solely on command logic and how you want your bot to be.

Comments URL: https://news.ycombinator.com/item?id=49727647

Points: 1

# Comments: 0

Article URL: https://github.com/snellingio/system-one

Comments URL: https://news.ycombinator.com/item?id=49727643

Points: 1

# Comments: 0

Article URL: https://snagflat.com/

Comments URL: https://news.ycombinator.com/item?id=49727637

Points: 1

# Comments: 0