Ten GitHub Projects Built with Jev Decision Models

Agents make dozens of small decisions before producing useful output: which source to trust, which tool to call, what context to keep and when to stop. Jev brings those hidden choices into the open by turning them into typed decisions with probabilities, instead of relying on free-form text alone.

That makes it useful across browser agents, code reviewers, context-pruning tools, search workflows, routers and mobile automation projects. In this article, we’ll look at ten GitHub repositories that show practical ways to build with Jev.

What Is Jev?

Jev is TypeSafe’s System One decision model. You give context, called state, and questions with predefined output types. It returns decisions and probabilities rather than writing an answer in free text.

The API supports choices from a supplied list, scores and yes-or-no judgements. That makes it useful for routing, ranking, filtering, and checking progress. Your application still defines the available actions and executes them. A valid output type does not guarantee a correct decision.

1. Jev Ultrafast

Best for: Developers exploring browser agents with explicit action choices.

Jev Ultrafast turns the current page into an action table. Jev chooses an operation and an observed element ID, and the browser executes the action.

The project also uses a separate text model to write input for TYPE_TEXT. Jev handles the decision; another model handles text that cannot be selected from predefined options.

Useful features:

• Browser operations and targets are selected together.
• An inspector following the action loop.
• A recorded Google Flights example in the repository.

GitHub: Jev Ultrafast

2. Fast Jev Compaction

Best for: Developers trying to control the context used by coding agents.

Long tool outputs can crowd an agent’s conversation history. Fast Jev Compaction uses Jev’s judgements about tool calls and results to decide what to keep, truncate or drop. It retains the original user and assistant text instead of generating a replacement summary.

This connects to context engineering: decide what the next turn needs, then remove lower-value material. Pruning can still discard useful evidence, so inspect the retained history.

Useful features:

• A standalone npm package and Claude Code plugin.
• Preservation of recent and initial messages.
• Fallback to built-in summarisation in the plugin when pruning cannot be used.

GitHub: Fast Jev Compaction

3. SemIf

Best for: Developers who want to explore typed decisions with local open models.

SemIf reads probabilities for predefined options from an open model’s logits. The application receives a typed decision without asking the model to generate an answer and then repairing its JSON.

SemIf was formerly called OpenJev and now uses the repository path SemIf-OpenJev. It is independent of TypeSafe. It does not provide Jev’s weights or reproduce its undisclosed architecture and training.

Useful features:

• Direct scoring of supplied answer options.
• Backend options for CUDA, Apple Silicon and CPU/GGUF use.
• Examples for testing the decision interface locally.

GitHub: SemIf

4. Jev Trader

Best for: Studying how a typed decision feeds into an automated execution loop.

Jev Trader is an experimental bot for the Kuru MON-USDC order book on Monad. It asks for a buy-or-sell direction and uses that decision to place a limit quote. A dashboard streams the order-book events and bot activity.

You can inspect how the market state becomes a decision and then a quote. The repository does not establish that the strategy is profitable.

Useful features:

• A TypeScript bot built with Bun.
• Streaming events and monitoring dashboards.
• A dry-run mode with simulated fills.

GitHub: Jev Trader

5. Hermes Jev Skills

Best for: Adding decision steps to an existing agent workflow.

Hermes Jev Skills packages Jev for model routing, memory relevance, context compaction, skill selection and other agent choices. The primary LLM continues to write responses while Jev helps decide how the workflow should proceed.

The basic concepts behind agentic AI explain how these components fit together. You can add one decision skill without rebuilding the entire agent.

Useful features:

• Several decision skills with configurable profiles.
• Integrations for Hermes and coding agent workflows.
• Shadow routing that logs recommendations before enabling model changes.

GitHub: Hermes Jev Skills

6. Jev Review

Best for: Exploring a decision-based review layer for code changes.

Jev Review examines a Git diff or a selected source scope and uses Jev’s typed judgements to flag issues, classify their severity, and determine whether changes meet defined quality criteria before merging.

Useful features:

• Diff-level and file-level decision scopes.
• Severity classification for flagged issues.
• Configurable quality thresholds for pass or fail outcomes.

Which Repository Should You Start With?

The right starting point depends on what you want to learn. If you’re new to Jev, Jev Ultrafast offers a clear end-to-end example of how typed decisions drive browser actions. If context management is your concern, Fast Jev Compaction demonstrates how Jev judgements can replace heuristic pruning. Hermes Jev Skills is the most modular option if you want to slot decision steps into an existing agent without redesigning it. SemIf is the best choice if you need to work entirely with local open models. Each repository is narrow enough to read in an afternoon, making them practical starting points for understanding where typed decisions add the most value in an agent pipeline.

Frequently Asked Questions

Q: Is Jev a general-purpose language model? No. Jev is a decision model. It returns typed outputs such as choices, scores and yes-or-no judgements rather than generating free-form text.

Q: Can I use Jev with models other than those from TypeSafe? Yes. Several repositories in this list, including Hermes Jev Skills, show Jev operating alongside other LLMs that handle text generation while Jev handles decisions.

Q: Is SemIf the same as Jev? No. SemIf is an independent project that reads probabilities from open model logits. It does not use Jev’s weights or replicate its architecture.

Q: Does Jev guarantee correct decisions? No. A valid output type does not guarantee a correct decision. Your application is responsible for defining available actions and evaluating results.