JevControl
Imports agent traces from LangSmith, Langfuse or OpenTelemetry and tests whether a smaller decision model can make routine choices.
Overview
JevControl, part of the Nokast open-source community, helps you find out whether a smaller model could handle some of an AI agent's routine choices, such as should I search again, which tool should I use, or is there enough information to respond, while the main model keeps writing the final answer. You import a record of an agent run from LangSmith, Langfuse, OpenTelemetry or a plain log, and JevControl shows the run step by step and highlights steps that look like small decisions rather than answer writing.
You decide which choices to test, connect a small model (one already running, one downloaded from Hugging Face, or a paid API), and JevControl runs the task twice: once normally and once with the small model making only the chosen decisions. Results show differences in time, model calls, tokens and cost where available, whether the small model agreed with the original choices, quality on tasks if you provide scores, and a ready prompt if you want to make the change. It never changes your production agent, and importing and reviewing traces requires no model. The README's travel-agent example tests a search again, show options or ask a question choice; its demo numbers are illustrative.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Should the agent search again, show options, or ask a question? (example routine choice) | search again · show options · ask a question |
How it works
- Import an agent trace
- JevControl flags steps that look like small decisions
- User selects which to test and connects a small model
- Run both ways and compare time, cost, agreement and quality
Strengths
- No changes to production agents
- Trace review works without any model
- Side-by-side comparison with agreement measurement
Caveats
- Demo figures are illustrative, not measured
- Comparison runs need access to the models
Getting started
python -m venv .venv && . .venv/bin/activate && pip install -e ".[dev]", then try the sample.At a glance
- Type
- Project
- Status
- prototype
- Language
- Python
- Runtime
- Local web app / CLI
- Has code
- yes
Use cases and integrations
How curated lists describe it
Find out whether a smaller model could handle some of your AI agent’s routine choices.
hellogumbo/awesome-jev
Frequently asked questions
Which trace formats are supported?
LangSmith, Langfuse, OpenTelemetry exports and plain log files.
Will it change my agent?
No, it only analyses traces and runs side-by-side tests.
Does it need Jev specifically?
No, you can connect any small model for the tested choices.
Related projects
Local-first, outcome-aware model router for coding harnesses that picks a capability tier per task with Jev and learns from history.
ProjectprototypeRustRuns Claude Code on Opus 5.5 with Jev re-picking reasoning effort after every tool batch, via per-message effort so the cache survives.
ProxybetaTypeScriptClaude Code plugin that routes every subagent spawn through Jev to pick its model, effort and tool profile.
ProjectprototypeLocal-first routing engine for long-running agent tasks that uses a Jev decider to pick a model group per turn.
PluginproductionRust, TypeScriptOpen-source AI gateway for 100+ LLMs that can use Jev to classify requests for complexity routing and to compact tool results.
ProxyproductionPythonOpen-source workspace for running and reviewing AI coding agents across devices, with an optional Jev-based automatic model router.
ProjectproductionTypeScript