jev-codex-router vs Poolside AI

Poolside's capability story is the models themselves: Laguna XS 2.1 (33B params, 3B active, 256K context) built to run on-device for lightweight work, and Laguna S 2.1 (118B params, 8B active, 1M context) aimed at frontier-class long-horizon reasoning. The sparse-active design is the efficiency argument. Around them sits the Poolside Platform: single- and multi-agent orchestration with planning and tool use, sandboxed execution for generated code, a desktop app and CLI for agentic sessions, model selection and tool grouping in the client, connectors to repos, databases and warehouses, RBAC for both humans and agents, and end-to-end trace observability. OpenRouter and Vercel AI Gateway are the listed integrations.

jev-codex-router solves a different problem entirely: it sits between Codex and the backend and picks model plus thinking-effort depth per call, not per session, via four Choice questions (Astra policy, capability tier, effort depth, route lease). Routing continues across tool-call continuations, not just turn one, and the mandatory Astra policy handles architecture and risk review. Responses pass through as a verbatim SSE relay — no format conversion. Two projections keep the decision state and the model's canonical replay separate; failures fail open, a sentinel file kills Jev instantly, and quota fallback only triggers on observed native quota exhaustion. Decisions are logged locally and never published. It integrates with Codex and LiteLLM. One is a model-and-governance stack; the other is a cost-routing middleware layer.

Poolside is pricing_type: contact — there is no published per-seat number, no self-serve free trial, and no instant sign-up. The page explicitly flags that buyers needing a publicly listed per-seat price before budget approval are not a fit. That means the real cost isn't just license fees: it's procurement time, security review, and whatever infrastructure you run the open weights on, whether that's on-prem, air-gapped, or multi-cloud. The trade is that you own and can fine-tune the weights instead of renting access. Worth noting from the news cycle: Bloomberg reported on 2026-08-22 that Nvidia will pay Poolside a $6B license, per an anonymous source — a signal of commercial validation, not a price you can plan against.

jev-codex-router is free. It's an MIT monorepo on GitHub (0xNatoshi/jev-codex-router) with an install.sh, markdown docs, and a local server component. The cost is your time: setup, maintenance, and comfort working from a README. There's no vendor SLA, no compliance certification, no published release notes, and no measured quota-savings number you can check before adopting — the project publishes a backtest protocol and limitations instead. If you were hoping to compare dollars, there's nothing to compare: one is a negotiated enterprise contract, the other is zero dollars plus engineering hours.