Felix Ortiz: Tess, hey — you see the date on this thing when you were prepping?
Tess Hollis: The four-day gap? Yeah. That's the one.
Felix Ortiz: Right, so — Claude Code Router v3.1.3 drops October 9th, and then four days later there's this analysis asking when local routers stop being enough. Four days. I don't know if that's perfectly timed cynicism or if it's accidentally showing us how fast this category is actually moving, but either way I can't stop thinking about it.
Tess Hollis: And that's the thing — Anthropic frames the whole local-router-to-gateway transition as a governance decision, not a performance one. So it's not 'your router is slow,' it's 'your router doesn't know who's using it or why.' That's a different problem.
Felix Ortiz: Yeah and — okay, so today we're getting into exactly that. When do local routers actually stop being enough? Is this a real scaling wall, or is it a category being narrated into existence the week a new router ships?
Tess Hollis: And what I want to sit with is whether 'enough' is even the right frame. Because the analysis assumes local routers were sufficient at some earlier point — and I'm not sure that's true.
Felix Ortiz: Oh — wait, that's a real question. Like, sufficient for who? One developer running Claude Code on a desktop? Sure. But past that?
Tess Hollis: Past that is exactly where it gets slippery.
Felix Ortiz: Okay but past that is — yeah, that's the thing I keep getting stuck on. Like, picture a developer on a Saturday afternoon, home office, running Claude Code Router on their laptop. That setup is genuinely complete. Everything they need is right there.
Tess Hollis: It's a smart power strip.
Felix Ortiz: Wait — say more.
Tess Hollis: Claude Code Router is basically a smart power strip. It knows which device to plug into which outlet — Claude, GPT, whatever. For one developer, that's the whole job. But the moment your engineering manager asks 'how much did your team spend on Claude calls last sprint,' the local router just... has no answer. There's no answer. It lives on one machine, it routes one developer's requests, it has no concept of other users even existing.
Felix Ortiz: Right, and — oh, that's actually where the analogy really lands. Because a smart power strip can't track which person in the building used how much electricity, can't cut off one floor if they go over budget. You need a breaker panel with a whole building-management system behind it. That's not a smarter power strip, that's a different category of thing.
Tess Hollis: And Anthropic's own documentation calls that out explicitly — it's a governance gap, not a throughput gap. The router isn't slow. It just has no centralized authentication, no usage attribution, no per-team budgets, no rate limiting, no audit logs.
Felix Ortiz: None of those. Zero. Which — I mean, for one dev that list doesn't matter at all, but the moment a second team joins, every item on that list is a live problem, yeah?
Tess Hollis: They don't just matter — they compound. And that's the part worth sitting with, because the gaps don't announce themselves. You just slowly lose visibility into what's happening across your org. And that's where the real question starts.
Felix Ortiz: And that's where the gateway options actually matter, because — okay, so you've decided you need one. Now what? LiteLLM is the obvious first reach. Open-source, MIT licensed, self-hosted, works with 100-plus providers. That's a genuinely great starting point.
Tess Hollis: Self-hosted being the operative word.
Felix Ortiz: Right, and — yeah, that's where I keep tripping. Because LiteLLM gives you SSO, SCIM, OIDC auth, audit logs in the enterprise tier, which sounds like it solves the whole governance list. But patching, scaling, wiring in your observability stack — that's all yours. Every bit of it.
Tess Hollis: Which is fine if you have a platform team that wants to own that surface. What's the version of this you'd only tell a friend? Smaller teams are massively underestimating that operational cost.
Felix Ortiz: Yeah, I think that's real. And then you look at Portkey — 250-plus LLMs, managed SaaS tier, private cloud option. Operationally it's just... lighter to carry. But then there's the acquisition.
Tess Hollis: Palo Alto Networks buying Portkey is the buried trap in this whole decision. The open-source repo is MIT licensed today — v1.15.2, January 12th, 2026. But 'today' is doing a lot of work in that sentence when the acquirer is a security company with its own product portfolio.
Felix Ortiz: Oh — so practitioners aren't just being paranoid about this?
Tess Hollis: They're actively flagging it before standardizing on the repo. Which is the right instinct. And then OpenRouter is the other end of that spectrum — zero infrastructure, nothing to host — but no audit artifacts, no SLA for security incident notification. It just disqualifies itself from any compliance deployment.
Felix Ortiz: Kong is the one that almost sidesteps all of this — if you're already running Kong, the AI plugins on Envoy are a natural extension. But if you're not, you're not adopting a gateway, you're adopting a whole platform to get a gateway. And the compliance angle — that's actually where this gets structurally worse, and we'll get into why entire product categories get disqualified before you even configure anything.
Tess Hollis: And that disqualification isn't a configuration dial you can turn. That's what I keep wanting to name precisely. Ignacio Peyrano published a paper in April — April 28th, 2026 — formalizing exactly this. Three-layer Zero-Trust model: Semantic Firewall, Tool-Level RBAC, Cryptographic Human-in-the-Loop controls. That's not a vendor blog post. That's academic architecture work on what an AI gateway actually has to be.
Felix Ortiz: Wait — like, peer-reviewed formalized architecture? For gateways specifically?
Tess Hollis: For semantic gateways in AI-native enterprise systems. And here's what that signals — the compliance bar is moving faster than the products. The research is already ahead of what any of these vendors are shipping.
Felix Ortiz: Oh. So today's compliant option might — wait, no, that's actually kind of alarming. Because if the formal architecture says you need cryptographic human-in-the-loop controls, and none of the current products fully implement that three-layer model yet, then what are teams even certifying against right now?
Tess Hollis: That's the unsettling part. And OpenRouter is the sharpest case. It's not that it's misconfigured — it publishes no audit artifacts. No SLA for security incident notification. That's categorical. You cannot put it in scope for SOC 2 work or anything FedRAMP-adjacent. Full stop.
Felix Ortiz: Not 'tricky to configure' — just structurally out. That's a different kind of problem than I was framing it.
Tess Hollis: And then Portkey sits in this genuinely weird position. Because Palo Alto Networks buying them could be a compliance asset — enterprise security pedigree, credibility in regulated environments. Or it could mean the open-source MIT-licensed repo you standardized on starts drifting toward a product that serves Palo Alto's portfolio, not yours. Practitioners haven't resolved which yet.
Felix Ortiz: Right, and — yeah, that's the one I can't get a clean read on. Like, is the acquisition the thing that finally gets Portkey into a FedRAMP conversation, or is it the thing that makes the roadmap opaque right when you need it most?
Tess Hollis: And that's the actual question worth sitting with. If the Peyrano architecture is where compliance requirements are heading — Semantic Firewall, cryptographic controls — then the gateway you pick today has to survive a compliance bar that doesn't fully exist yet. That's not an infrastructure decision anymore.
Felix Ortiz: Yeah, I don't have a clean answer. Like, in 18 months, does the October 13th analysis look like real foresight, or does it just look like the timestamp when someone finally wrote down what vendors were already selling? The Claude Code Router v3.1.3 release being four days earlier — that might matter less as a signal and more as... a marker. The moment the category stopped being theoretical.
Tess Hollis: And whoever consolidates — LiteLLM, Portkey, OpenRouter — someone's on the wrong end of a roadmap decision they didn't see coming. That's the part I can't resolve. MIT licensed today doesn't mean MIT licensed in 18 months if Palo Alto Networks has a different idea about what Portkey should be.
Felix Ortiz: Right, and — the 'model swap as config change' argument is genuinely compelling. Architecturally it's real. But it's also exactly what every gateway vendor leads with. So I don't know how to hold both of those things at once.
Tess Hollis: You probably can't. That's the honest place.
Felix Ortiz: Yeah. I think that's actually — yeah. We spent the whole conversation trying to find the clean inflection and I'm not sure it exists yet.
Tess Hollis: Good conversation to not finish.
