Office Hours · 22 October 2026

OpenLint Office Hours #4

Summary

The fourth OpenLint office hours focused on the formal specification and on how OpenLint can help teams distribute their rules.

Specification. Joost Farla’s rewrite of the specification as a formal draft, built with Bikeshed, was ready to merge. It includes tooling that compiles the spec and runs its conformance tests against the current implementation. It also records, inline, the many places where behavior was reverse-engineered from Spectral rather than documented, such as multi-document YAML files, where only the first document is linted. The group agreed to merge it as a starting point and work through the open questions as smaller issues and discussions. It’s now published at spec.openlint.org.

Toolbox. Phil and Jakub weren’t on the call, so the npm release status carries over to next week. Miguel Quintero is working on TypeScript 7 and an MCP server: read-only to start, and designed MCP-first rather than generated from an API. Miriam Greis described companies building agent skills that run linting and open pull requests for easy fixes, and the push at Nordic APIs to make existing APIs “AI ready”. Kin connected this to the need many enterprises have for distributing, versioning and evolving their rules, beyond Git and an editor plugin.

Research and launch. Rules research shows most teams use the default ruleset unchanged, so OpenLint’s defaults matter. Kin is studying the Dutch government’s approach as a blueprint, aiming for a formal OpenLint announcement at API Days Paris, alongside its Open Source Europe government track.

Funding. $1,000 is in the Open Collective account, with European Commission grants to pursue now that Open Source Europe is the fiscal host.

Also discussed: using JSON Schema plus OpenLint rules to constrain the inputs and outputs of open AI models, for example FHIR, ISO 20022 and PSD2, as a side experiment. Future calls move an hour earlier.

Agenda and discussion on GitHub · Watch on YouTube

Transcript

00:00:00Before the start: Personal catch-up while people join. Then talk turns to an event Kin attended in New York the day before, where the common story was "we're just starting to get a handle on API sprawl, and now we've got MCP sprawl, agent skill sprawl, and everybody vibe coding".

00:02:25Kin Lane: Welcome, everybody. I'll share the agenda, and I'll fully admit I don't have my act together this week after Stockholm. I'm still jet-lagged, and I had another event here in New York yesterday, so I honestly haven't done a whole lot. But it looks like Joost has, and I see Phil and others have. Here's the agenda. I published the office hours from last time, but that's about as far as I got. Joost, what have you got cooking on the specification?

00:03:14Joost Farla: Following the previous session, I've been working on it, and I created a first pull request with the specification. It's not complete yet, and it's not perfect, but I thought it would be good to already have something published to look at, which can help the conversations and the toolkit. The pull request has a rewritten specification, plus tooling needed to compile the specification and to help with checks and releases. There are a lot of open issues in it: terms we have to discuss, ambiguities, and differences between the specification and the implementation. I think it would already be good to have this online and proceed from there. I plan to do more iterations to make it more established and more accurate, but I think it's a solid base for the current state of the project.

00:05:16Kin Lane: Thank you for the work. It's good to see, and much more rigorous than the first pass I did, both for the site and for fleshing out the issues with the spec. It wasn't a spec before, so a lot of things will have to be tightened down and cleaned up.

00:05:42Joost Farla: Yes. There are also many requirements that aren't in Spectral's documentation but are reverse-engineered from the implementation. The tooling also has a check that runs the conformance tests against the current implementations, so we can already see whether they're consistent. Right now we reverse the specification from the implementation, and verify that the implementation complies; later it should go the other way around, with the specification leading.

00:06:37Kin Lane: We'll have to triage the severity of the things we reverse-engineer, and decide what we can implement without major changes: find the sweet spot between the tooling and the spec that's good enough to go out the door, without introducing anything breaking or drastic.

00:07:07Joost Farla: I already ran into a lot of undocumented behavior that the current implementation accepts, and that's debatable. For example, a YAML file with multiple documents separated by ---: both Spectral and OpenLint accept it and just take the first document. I'm not sure we want that. But that's something we have to discuss: how we deal with these kinds of issues.

00:08:03An attendee: The issues you talked about: are they in the files?

00:08:10Joost Farla: Yes, they're in the specification, inline.

00:08:15An attendee: There are a lot of issues and concerns there. Should we turn them into real issues? How are you planning to work on them, since different people work on different things?

00:08:29Joost Farla: Some are already in current discussions, and where they are, I've linked to them. For others, I think we need new discussions or issues to find solutions.

00:08:47Kin Lane: I'll prioritize time today, tomorrow and through the weekend to go through it and familiarize myself, then suggest opening new issues where needed, and get my head around the existing ones. This is pretty critical. We don't have Jakub or Phil on the call for the code side, but I think this will have to move forward bit by bit. Go ahead, Mike.

00:09:32Mike Kistler: If we can merge it, it becomes available on GitHub Pages.

00:09:38Kin Lane: Right, on Pages, and not just on the custom domain. If you're fine with it, Joost, I'm fine with merging and publishing it with the domain, so we can start working from it. Or would you rather keep it as a PR for now?

00:09:59Joost Farla: I'd prefer to merge it, so we have a starting point, and then we can do smaller iterations in separate issues and discussions.

00:10:11Kin Lane: That's what I'd rather do too: get it up, because we're trying to get open for business, and the communication around it is more critical. Then we can break things up into smaller discussions, issues and pull requests. Mike?

00:10:30Mike Kistler: I was looking at the PR, and the specification moved out of the Markdown file into a different file. What's the .bs format?

00:10:41Joost Farla: It's Bikeshed, a tool for the whole specification lifecycle, and it has its own file extension. It's for building specifications: references to normative documents, consistency checks and so on. It's like ReSpec on steroids.

00:11:08Kin Lane: Thanks, Joost. I'm going to merge this after the call and set up the subdomain, because having it up on GitHub Pages, for communication purposes, is pretty key. What I love about these standards calls is how much I learn sitting in on them; that's the first time I've heard of Bikeshed. I'll set up the subdomain right after this call, prioritize it, and then work my way through the issues.

I don't think we have an npm package yet, last I checked.

00:13:01Miguel Quintero: I don't think so. I wanted to ask if you can add me to the npm organization. I'm logged in, but I don't see the organization on my npm profile.

00:13:15Kin Lane: I neglected that after the last call, so I'll add you.

00:13:20Miguel Quintero: I was trying to ping Phil to see what he's working on, so I could go ahead. I'm doing two things: TypeScript 7, and the MCP server. I opened an issue for the MCP server to throw out some ideas, so I can then open PRs, and I'm looking for feedback on those ideas. The original issue was MCP *and* API; I'm taking a different approach, a separate pathway: MCP design first, rather than API to MCP.

00:14:29Kin Lane: I see where that goes: MCP design first, versus API to MCP.

00:14:36Miguel Quintero: Exactly, that's the approach I'm taking. Right now it's read-only. Say you're an API engineer at a company, or the company itself, trying to pull up their API: "What do my specs look like against the standard?" It doesn't edit anything. I just put some ideas out there, and I'm hoping we get some feedback.

00:15:28Miriam Greis: What's also interesting: one company I know created their own skill. They already have many APIs, and they created a skill to set things up correctly, to make sure it runs correctly, and to fix things already: adding PRs for the easy findings, and flagging the more difficult ones. So I think this is getting more relevant. At Nordic APIs this was a topic too, with multiple companies trying to make their APIs "AI ready". How do I get quality into existing APIs? Detecting changes, which rules to impose, breaking changes and so on. We could support that with an AI solution as well. I think that's a really good thing.

00:16:41Miguel Quintero: Exactly. Start with something simple, even simpler than an API: compliance, and then use it with skills.

00:17:27Kin Lane: Sounds great, excited to see it. Before I add to that: Joost, on the spec, do you need anything beyond me merging, setting up the subdomain, and giving wider feedback?

00:17:46Joost Farla: No, I'll just continue iterating on it. I'm curious what you think once you've been through it.

00:17:59Kin Lane: That's my priority for today and tomorrow, so I'll give you feedback by next week.

On what you're working on, Miguel, and what you said, Miriam: I've had a few enterprise customers really looking for this as the next level of distribution to teams. A lot of companies have implemented Spectral rules, but other than having them in Git and a VS Code plugin, they didn't really have any distribution policy or approach: how rules are versioned, updated and evolved. One of the things that triggered me forking Spectral was that I was building an API and MCP server for API Evangelist. I was going to offer it as an open-source server with a support package around it: I'd help you configure it with your rules and style guides, and help you build a set of skills for your process. The server itself is open source, you run it and you own it, and you'd pay me to help dial it in for your teams, as a distribution mechanism.

One last thing: I've been joining the skills-over-MCP standard calls with Anthropic over the last couple of weeks. Having your skills deployed as a formal primitive in your MCP server, to help stabilize skills distribution, is brand new; very few people have adopted it. It's one of the things I'd like to explore, and I'll leave feedback on your issue. When you first set up an MCP server, how do you configure it for your organization: with your rules, your style guides, your skills, how your teams and orgs work? I think that's going to be really, really useful.

00:20:25Miguel Quintero: Yeah, exactly. Lean on the defaults as much as possible, and go from there.

00:20:51Kin Lane: What else? We don't have Phil and Jakub to talk about the core product and the package, so I'll try to pay attention to what's happening asynchronously, and we can come back to it next week. What's top of mind for anybody? I have a couple of other items, but I want to open it up first.

I'm still doing more rules research based on the rules that are out there, and spending more time on the Dutch government's approach to rules. I'm going to push on Phil, and we'll see if we can pull it off, but I'd like to announce OpenLint formally at **API Days Paris**. On the second and third days there's an Open Source Europe part that brings together European governments to showcase what they're doing with APIs. I want to use API Days Paris plus that to showcase OpenLint. So I'm studying how the Dutch government does what you do, as a blueprint: your rulesets, your rules distribution, your guidance. There's a lot of good stuff to use as a blueprint.

One thing I found in the rules research and the pipeline research is that **most people just use the default ruleset out of the box**. They don't make any changes. That's going to have a big impact on what OpenLint's default settings are, and how we let people configure them; it's going to have to be very smart.

On document types, I want a really robust list to bring to the table. Not that we'll push them all onto the roadmap right away, but it's an important way, around the announcement, to make people aware that this is about more than just OpenAPI.

Money: we've got $1,000 in the account. My goal is to start begging and harassing other companies for money, so we can pay Phil, Miguel, Jakub, you, and others doing the work. Pavel shared a couple of links to European Commission grants we can submit to, now that we have a financial partner in Open Source Europe. I want to chase sovereign money and matching funds, to pay everybody for project, tooling and standards work, with little micro-project funds, so we're not all just doing it for free.

Oh, and I was supposed to move this call an hour earlier. I just didn't have my act together this week; I apologize. As soon as we're done, I'll merge your pull request, Joost, set up the spec site, and move next week and future calls an hour earlier.

Anything else you want to talk about? We're making good progress, and I'm really excited about the potential. I feel like OpenLint, Arazzo and Overlays are the tip of bringing new energy to the older, ongoing conversations and hard work we've all been doing over the years.

One side thing: DeepSeek launched an infrastructure harness around their open model, and one package does schema input and schema output, so you can restrict what goes in and out of the model. If you think about what Jev [?] is, it's a decision out: a yes, a no, and a probability, so it's a schema restriction. I really believe JSON Schema plus OpenLint, as a linting mechanism in and out of open models, has a lot of legs. I don't really know the details, to be honest, but I'm going to run experiments on AWS with the DeepSeek stuff, with the FHIR specification, PSD2 banking and healthcare in and out. When you start talking about industry-level and sovereign-level rules for inputs and outputs, I think there's interesting work there. It's not directly related to what we're building, so I don't want to muddy the water, but I figured I'd share.

00:28:22Miguel Quintero: How do you feel about using OpenLint with those decision models, to be slightly more deterministic? Pointing it at the model and saying, "Help me lint this; what are the problems with my API?"

00:29:35Kin Lane: To be honest, I don't fully understand it; models aren't my specialty, and neither is the infrastructure around each model. I've had a taste with Claude, which does a lot around the inputs and outputs of its models, and I've recreated a lot of my tasks with an open-source model I run locally, and learned fast that they've done a lot of magic there. *(To Joost, leaving:)* Thanks for your PR, Joost. Next week will be earlier, I promise.

From my vantage point, using JSON Schema plus Spectral to govern APIs, they work together: JSON Schema is foundational (here's what should go into an API, here's what should come out), and rules work somewhat like overlays on top, adding completion rules with regex, functions and so on. And the rules are domain-based: a core JSON Schema everyone agrees on, and then, per domain, we decorate, overlay and evolve the governance with different sets of rules. This infrastructure around model inputs and outputs is kind of like a gateway to me. A classic API gateway applies JSON Schema to what it pulls in and what it publishes, and it has to be valid. Runtime isn't always the best place to do it, but in some situations it works. The decision-style approach limits tokens and steers the non-determinism towards determinism, but it feels like one casino game. Depending on whether you're in healthcare, and what jurisdiction or country you're in, you might want different schemas and different sets of rules.

00:33:17Miguel Quintero: That's a very good point, especially for very large enterprises with many different businesses. Same company, but different priorities: rules here, rules there, schemas here, schemas there. They're fragmented within the same company. It's really hard.

00:33:58Kin Lane: Everyone copying that decision approach takes one of two paths, but they all share a schema: yes, no, then a probability. That's a schema on the output of the harness. So that's the core schema; could you then do a rules overlay, so different teams choose their rules about when a yes or no is valid, and when a probability is acceptable? Domain-based or team-based rules on top of a shared schema. That's one way OpenLint could fit, on the output. What you'd do on the input, I don't know. And there are costs: are we limiting the data, versus the probability? The reasons you'd want schemas or rules would vary by industry and by who you are. I want to build a playground to understand it, because I know schemas and rules, but not models: strict schema in, strict schema out, rules overlaid on both, and then swap the models and see what you get. You're still relying on a model, but if you run a lot of tests against different models over time, maybe you can build some confidence that it works.

00:36:34Miguel Quintero: You're going in the right direction. *(Partly inaudible.)* With the frontier models, you don't just hand them the problem; you give them the experiment and your parameters and constraints, and they do the research. With these decision models, you'd better have those questions ready, and know what scores they'll give you.

00:37:48Kin Lane: That's what I want to do with other schemas. The decision output is just one schema. I want the FHIR specification, or ISO 20022 for payments. Can I take a PDF or a blob of text that's a patient record, put it in, and tell the model it's constrained coming out: it has to be FHIR, which version, with a set of rules that validate not just the schema, but that the data matches certain criteria? Can you squeeze a more trustworthy response out of a model in a healthcare scenario that way? I still feel there's non-determinism at the core. But try it with FHIR, with ISO, with PSD2, with different models, and run tests: is it acceptable enough? This is where sovereignty and jurisdiction come in: in the US that might be good enough, and in Europe it won't be.

Interesting conversation, probably for another time. I'll nudge Phil and Jakub on where we are with the core packages. Joost's work is the most progressive this week. I'll get it published and send a note to everybody once it's up, so we can start going through it, lighting up actual issues, and aligning it with what Phil, Jakub and you, Miguel, are doing. I'll spend time on your MCP issue, Miguel, and layer on the work and thoughts I already had, because MCP and skills is a really interesting one that a lot of people care about.

00:40:21Miguel Quintero: Go at it; I just put a baseline there. And Miriam, I'm hoping you can weigh in when you have time, especially from the governance adoption perspective.

00:40:48Kin Lane: Thanks, y'all. See you next week.