Spotlight: Septor Labs — a router that bills agents by the hour
Septor Labs sells the thing you'd build yourself if you had a spare month: one endpoint that talks OpenAI's API shape but sits in front of a hundred-odd models, open-weight, open-source, and frontier, with routing, caching, and observability baked in.
The problem it's solving is familiar to anyone who's shipped past the demo stage. You start on one model, then you want a cheaper one for simple requests, a stronger one for hard ones, a fallback when your primary provider has a bad night, and some way to know what any of that actually cost you before the invoice arrives. Normally that means stitching together SDKs, writing your own router, and bolting on logging after the fact. Septor's pitch is that you point your existing OpenAI client at their base URL and get all of that for free: adaptive routing with latency-aware fallbacks, a semantic cache, and per-request traces with cost breakdowns.
The smart decision here is the drop-in compatibility. There's no new SDK to learn and no re-architecting, you just change a URL and a key, which means the cost of trying it is close to zero and the cost of leaving is just as low. That's a good trust signal from a young infra company: they're not trying to lock you in, they're betting that once you see the routing and the observability pane, you won't want to go back to a single hardcoded model.
The more interesting bet, though, is the agents. OpenClaw is a headless browser agent that runs sandboxed Chromium and reasons over the rendered DOM, so it can fill forms, scrape behind auth, and take screenshots. Hermes is a persistent async agent with state on a mounted volume and a webhook for events, built to survive restarts and run for hours. Both are billed by the hour rather than by seat or by request, with explicit CPU and memory caps and what they call a fail-closed credit gate, meaning the agent stops instead of quietly running up a bill you find out about later. Anyone who has watched a runaway agent loop burn through a budget overnight will recognize why that's the feature to lead with, not an afterthought.
Pricing is tiered and legible: a free tier with $5 in credits and 50 requests plus 200K tokens a day (no card required), then $20, $50, and $100 monthly plans that scale up pipeline counts and unlock more of the open-source catalog, with the top tier including frontier models. The "adapters, not rebuilds" idea, bolting on RAG, web search, or PII redaction via a single header, is the other quietly useful piece, since it means the same API call can gain retrieval or guardrails without a new integration.
This is a good fit if you're already calling models through an OpenAI-compatible client and want routing, caching, and cost visibility without owning that infrastructure, or if you're experimenting across multiple model providers and want one invoice and one trace view instead of five dashboards. It's also a reasonable place to prototype an agent that needs to browse or run long background tasks without standing up your own sandboxed browser infra.
It's probably not for you if you need a hard SLA from a single frontier provider and don't want a routing layer between you and it, or if you're deep enough into one model's specific tooling (fine-tunes, batch APIs, provider-specific features) that an abstraction layer costs you more than it saves. And because this is a new company without a published enterprise story or funding details, I'd be cautious putting mission-critical, compliance-heavy workloads through it on day one.
Where this goes next, in my read, depends on whether the routing and observability stay genuinely provider-agnostic as the model list grows, or whether Septor ends up nudging traffic toward whichever models are cheapest for them to serve. The agents-by-the-hour model is the part I'd watch closest: if OpenClaw and Hermes work as advertised, that's a distinct wedge in a market that's mostly competing on model selection alone.
Try Septor Labs: septorlabs.com
See the launch: Septor Labs on welaunch.sh