SpacetimeDB: The Database That Wants to Kill Your Backend

Good morning and welcome to the Builder's Briefing for February 27th, 2026. I'm Alex, joined as always by Sam. We've got a packed one today — a database that wants to replace your entire backend, Anthropic walking back safety promises, Google API keys suddenly becoming dangerous, and a wild amount of agent tooling all trending at once.

Yeah, seriously — when three separate LLM API gateways trend on the same day, that tells you something about where we are. Let's get into it.

Alright, the big story. SpacetimeDB hit the top of GitHub trending this week with over twenty-three hundred stars, and the pitch is pretty radical. Your database IS your server. You write stored procedures — they call them reducers — in Rust or C#, clients connect directly over WebSocket, and there's no REST API, no backend service, no ORM in between. It's a single binary that handles state, logic, and real-time sync.

Okay, so it's like if SQLite and a game server had a baby. And that's not just a cute analogy — there's literally an MMO called BitCraft already running on it. You define your schema as Rust structs, write your mutation logic, and clients automatically get updates when data they care about changes. No pub/sub to wire up, no cache invalidation headaches.

Right, and what's wild is this isn't an isolated trend anymore. Between SpacetimeDB, Convex, and Supabase's real-time features, the data layer is just absorbing application logic. If you're starting a new project with heavy real-time requirements — multiplayer, collaborative editing, that kind of thing — it's worth prototyping here before reaching for your usual backend stack.

I could see this becoming the default for collaborative features within six months. The developer experience of not maintaining an API layer is just... chef's kiss. That said, I'd want to understand the escape hatch story before going all-in on production.

Fair point. Okay, shifting to AI news. The one that raised eyebrows — CNN is reporting Anthropic has walked back a key safety commitment. If you're building on Claude's API, this matters because content filtering behavior could shift under you.

That's interesting because Anthropic's whole brand has been 'we're the safety-first company.' This reframes their competitive positioning against OpenAI pretty significantly. For builders, the practical takeaway is — don't hardcode assumptions about what the model will and won't do. Those guardrails can move.

Also in AI — DeepSeek V3 dropped a paper introducing something called LEANN that achieves ninety-seven percent storage savings for on-device RAG. If you're building local-first retrieval applications, this is huge. Running RAG on consumer hardware without cloud dependencies just got way more realistic.

And there's a company called SI Inc claiming they've shipped the first fully general computer action model — not just browser automation but full desktop interaction. If that holds up, it could replace those brittle Selenium and Playwright chains with a single model call. Big if true.

One more quick AI hit — researchers showed that LLMs can deanonymize users at scale by correlating writing patterns across platforms. If you're building anything with pseudonymous accounts, that's a threat model you need to think about now.

Yikes. That's the kind of capability that's both a moderation tool and a privacy nightmare depending on who's wielding it.

Alright, dev tools. This is where today gets really interesting because of the sheer volume of agent infrastructure that trended simultaneously. First up, cc-switch — a desktop app that lets you switch between Claude Code, Codex, OpenCode, and Gemini CLI from a single UI. Twenty-two hundred stars already.

That's hitting a nerve because everyone I know is juggling at least two or three coding agents right now. The fragmentation is real and painful. And alongside that, you've got claude-flow for multi-agent orchestration — deploying agent swarms with distributed coordination and RAG integration.

And then there are two separate LLM API gateway projects trending — one-api and sub2api — both solving unified routing across OpenAI, Claude, Gemini with key management and cost sharing. Plus Vercel Labs shipped just-bash, which lets agents write and execute bash directly as a runtime.

Just-bash is the most Vercel thing ever. Simple, a little dangerous, and probably exactly right for half the automation tasks people are overthinking with structured tool APIs. But the bigger picture here is that the unification layer for agents is being built in real-time, right in front of us.

Oh, and there's a practical tip — someone published an analysis showing you can slash MCP costs by routing through CLI tools instead of the full server protocol. If your Model Context Protocol bill is climbing, link in the briefing, worth implementing this week.

Security corner — and this one is urgent. Truffle Security found that Google API keys, which historically were low-risk, now grant access to Gemini models and can rack up charges. If you have exposed Google API keys in public repos or client-side code, rotate them now. Treat them like AWS keys.

Oh wow. That's a silent rule change that could burn a lot of people. How many tutorials and Stack Overflow answers have Google API keys sitting in client-side JavaScript because 'they're just API keys, they're fine'? That assumption just broke.

Exactly. Also worth noting — there's new research breaking Wi-Fi client isolation on enterprise networks, which matters if your threat model assumes network-level isolation for IoT or guest networks. And Vaultwarden, the Rust-based self-hosted Bitwarden alternative, is trending again with nine hundred stars. Still the gold standard if you want full control over credential management.

Love Vaultwarden. If you're running a team and don't want the Bitwarden cloud bill, just set it up. It's rock solid at this point.

So the big takeaway for today — the agent tooling layer is fragmenting fast, and builders are already shipping the unification layer. Three API gateways, a multi-agent orchestrator, and a coding agent switcher all trended on the same day. If you're building on AI agents, invest in provider abstraction now. The model you're using in March won't be the model you're using in June.

That maps to the Benedict Evans piece that also trended today — he's arguing OpenAI's moat is thinner than people think as models commoditize. The gateway projects exist precisely because smart builders already see that future coming.

And if you're starting a real-time collaborative feature, seriously give SpacetimeDB a look before defaulting to your usual stack. The database-as-server pattern is ready for production.

Agreed. Abstract your LLM providers, rethink your backend assumptions, and for the love of all that is holy, rotate your Google API keys.

That's the Builder's Briefing for February 27th. All the links are in the briefing notes. We'll be back tomorrow — until then, ship something great.