Thesis
jev-router makes model selection a per-turn routing layer around Claude Code and Codex, rather than a manual choice for each task.
Observed facts
The repository README describes jev-claude and jev-codex as launchers for the real upstream CLI, with model selection for a fresh user turn.
1It says setup requires Node.js 20.12 or later, a supported CLI, and a Jev API key.
1It documents a loopback proxy and a control for pausing routing by selecting a concrete model.
1
Mechanism
The router sends the prompt and model-tier information to Jev for a decision, then uses the selected tier while forwarding the CLI's existing authorization headers. The README says those headers are not read, stored, or modified.
Why now
The project documents support for both Claude Code and OpenAI Codex, making automatic tier selection a practical workflow option for users who otherwise choose models manually.
Implications
The first implication is that model choice becomes part of the agent workflow and can reduce manual switching between tiers.
A further implication is that the routing decision depends on an external decision service, so users should understand what task context is sent to it.
Evidence boundary
Evidence here is the public GitHub repository README and its descriptions of setup and behavior; it is not an independent test, security audit, or evaluation of routing outcomes.
Watch conditions
Check for independent measurements of cost and task quality across routed and manually selected models.
Check what prompt data reaches the Jev decision service and whether the project documents retention and privacy controls beyond the stated handling of CLI authorization headers.