Deno reimplemented Cloudflare Workers and Durable Objects as Apache-2.0 software you run on a $48 node

Ryan Dahl announced celld on August 5. It's a daemon built from V8, S3, SQLite, LTX and Tokio that runs the exact Cloudflare Workers and Durable Objects JavaScript APIs and configuration on hardware you own.

The architecture is the interesting part, not the API compatibility. Each object is its own SQLite database, addressed by name, replicated to an S3-compatible bucket you control. No control plane. No consensus protocol. Coordination happens entirely through the bucket. Writes are durable before acknowledgement, so RPO=0, and idle cells hibernate to near zero cost. Deno's number: an 8GB DigitalOcean node at $48/month holds up to 1,000 resident cells, which they position as an order of magnitude cheaper at scale.

It landed the same day Cloudflare open-sourced Cloudflare OS, its internal agent platform, built on those same primitives. I have no idea whether that's coincidence or a very fast response, and I'm not going to pretend I do.

Cloudflare OS is worth its own look regardless. Three parts: a browser-based Agent Workspace where agents run isolated code with company context and persistent state, a governance layer where service-specific "Gatekeeper" Workers mediate every resource request from a zero-permission baseline while tracking every resource an agent observes, and an app platform where agents build full-stack apps on Dynamic Workers and Durable Objects with SQLite. It ran internally at Cloudflare since May. No pricing announced.

That zero-permission baseline is the design decision to steal. Agents start with nothing and request capability, rather than inheriting your ambient credentials. It's the same instinct behind Cloudflare's separate Agent Access Model proposal, whose sharpest idea is capability ratcheting: once a capability is removed mid-task it requires fresh authorization, so trust can only narrow, never widen. That paper is explicitly proposed architecture rather than shipping product, built on OAuth Token Exchange (RFC 8693) and DPoP (RFC 9449), and it admits multiplayer access control is still unsolved.

But celld is the strategy story, because of what it's part of. In 48 hours: celld undercut Workers/Durable Objects. Superlog Responder shipped a free open-source production bug-fixing agent aimed at Datadog and Sentry. Neon and Castform published a post-trained 4B open-weights model matching GPT-5.6 Sol on agentic retrieval at 1/100th the cost. Mistral released Shieldstral 1.0 3B as Apache-2.0 weights against hosted moderation APIs. Prime Intellect open-sourced its MIT-licensed Prime Agent harness.

Five categories. No shared customers. Same move.

If your moat is runtime, retrieval, moderation, or harness scaffolding, it has a half-life measured in weeks right now. Price accordingly and architect accordingly. The Neon result in particular reframes how I'd build RAG today: treat the frontier model as the teacher, not the runtime. Post-train a small open model on your specific retrieval loop and the per-request economics change by two orders of magnitude. Their baseline was 10+ seconds and ~$0.03 per multi-turn agentic search against gpt-5.6-sol, because retrieval became a planning loop and every hop multiplies frontier token cost.