How do I install it, and is it in the VS Code marketplace?

The README documents no published listing. You run npm run build, then use Extensions: Install from VSIX and point VS Code at the repository directory, reloading when prompted. For development you can launch an Extension Development Host instead.

Is this a production tool or a research demo?

The README is explicit that it began as a proof of concept, a research experiment showing that an LLM given runtime context debugs better than one given source alone, and that it is not a supported product.

Which languages can it actually debug?

The documented path is Node.js: the diagram shows a Node.js debug adapter executing a Node.js application. Nothing in the README covers other runtimes or debug adapters, so treat Node as the only supported target.

Do I need both an OpenAI key and a Jev key?

You copy api.env.example to api.env and add OPENAI_API_KEY and TYPESAFE_API_KEY. Either brain can instead point at the Vercel AI Gateway via llmProvider or jevProvider, and the llmDebugger.strategy setting decides which one routes each step.

What makes the coupon-stacking example notable?

Told only that coupon percent applies before fixed produced 1450 instead of 1500, it placed its own breakpoints, stepped down coupons.js line by line, and named the offending line. It did that in 26 debugger actions and 27 fast checks, with only 2 generation calls, in about 17 seconds.

What do the two benchmark runs show about routing?

With gpt-5.4-nano, Jev routing solved 13 of 13 in a 12.8-second median run at $0.055, while letting the model route itself solved 11 of 13 in 92.5 seconds at $0.243. With gpt-5.6-sol at low reasoning, Jev also solved 13 of 13 at $0.480 against the model routing itself at $1.086, so Jev saved money with a frontier model and saved the task with a cheap one.

CommunityJev Tools & CLIs

LLM Debugger Vscode Extension

Q: How do Jev and the generation model split the work?

Jev routes every paused step and says whether the evidence is enough, from the stack, locals, nearby source, program output, and what has been tried. The generation model is woken only for the opening hypothesis, a breakpoint location, an expression to evaluate, and the closing root cause and patch. Deterministic code owns the step and wall-clock budgets, breakpoint validation, stall detection, and the evidence requirement.

A VSCode extension that debugs by driving the real debugger: Jev routes each paused step as a typed decision, and a generation model is called only when a breakpoint location, an expression, or a patch is needed.

See highlights

Open github.com

Stars: 361
Language: TypeScript
License: MIT
Published by: Community
Author: mohsen1
Last verified: Oct 2026

mohsen1/llm-debugger-vscode-extension

User repository on GitHub

View on GitHub

VSCode extension that demonstrates the use of large language models (LLMs) for active debugging of programs

Forks: 24
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.

About this project

A VSCode extension that finds bugs by running the code — driving the real debugger, setting breakpoints, stepping, and reading live values — rather than reading the source and guessing. Each step's "what next?" is answered by Jev, a typed System One model, for a fraction of a generation call. The generation model is woken only for what a classifier cannot produce: the opening hypothesis, a breakpoint location, an expression, and the final fix.

Quoted from the project's README. Read the full README.

Tagscommunity jevbest

Works forJev for Classification

Used inJev for Software & DevTools

Languages: TypeScript

Highlights

Jev answers what the debugger should do next from the paused state; the generation model handles locations, expressions, and the patch.
Deterministic code owns the step budget, wall-clock budget, breakpoint validation, stall detection, and the evidence requirement.
A verdict that has not read a runtime value is rejected, so a run cannot end by pattern-matching the error message.
The 13-task benchmark applies each proposed patch and re-runs the program; per-run tables live in benchmark/results.md.

Watch out

The README says this began as a proof of concept, a research experiment, and is not a supported product; the 13-task benchmark is built by the project itself. Installation is from a locally built VSIX rather than a marketplace listing, and the architecture diagram shows a Node.js debug adapter, so other languages are not documented. Copy reviewed against the repository README on 2026-10-04, not from running it.

Questions about this entry

Short answers drawn from the linked project or page.

More Jev resources like this

crush-monitor

288

A local React and Express web app that takes a pasted two-person chat transcript and annotates each message with emotion and intent probabilities plus a grade for your own replies. Chinese interface, your own API key.

Jev Tools & CLIs#community#jevbest

Community

FerryCorleone/crush-monitorTypeScript

OmniStudio

197

A Chinese-first desktop workbench for local models whose JEV page runs typed noul, choice and score questions against a local engine or a cloud key, alongside a gateway, RAG knowledge base and shared memory.

Jev Tools & CLIs#community#jevbest

Community

kunpengtalk/OmniStudioTypeScript

prism-liquidity-agent

124

Prism is an autonomous liquidity agent for Solana concentrated liquidity pools, currently Meteora DLMM. Every ten minutes it recalls past outcomes, reads pool state, scores volume authenticity and decides HOLD, REBALANCE, EXIT or ENTER behind a risk gate.

Jev Tools & CLIs#community#jevbest

Community

irfndi/prism-liquidity-agentTypeScript

jev-poly-crypto-demo

Research dashboard for Polymarket BTC five-minute Up/Down rounds that feeds market state to Jev class scores and paper-trades the edge.

Jev Tools & CLIs#community#jevbest#catalog-import

Community

frankda/jev-poly-crypto-demoTypeScript

jev-recall

Python memory filter that judges every stored memory against the current request with one calibrated yes/no per memory in a single Jev request, keeping everything above a threshold.

Jev Tools & CLIs#community#jevbest#catalog-import

Community

samdotmak/jev-recallTypeScript

sharp

Browser extension that filters your X timeline with a model you choose and pay for, using plain-language rules plus per-author and blocked-word controls.

Jev Tools & CLIs#community#jevbest#catalog-import

Community

tshmieldev/sharpTypeScript

From the Jev guides

Original write-ups on what this entry is for.

Jev on the Gateways: Vercel, Workers AI, and OpenRouter

Oct 12, 2026

The request shape barely changes across surfaces; the surrounding contract does. When to call the API directly and when a gateway, a Worker binding, or a router entry earns its extra hop.

Jev Tools & CLIs#gateway#vercel#cloudflare

Read article

Exporting Decision Models: GGUF, ONNX, and Device Runtimes

Oct 10, 2026

Taking a decision checkpoint out of Python: what the GGUF and browser WebGPU exports preserve, what they change, and the parity evidence the model cards publish.

Jev Tools & CLIs#open-models#local#onnx

Read article

Records at Rest: Documents, Inboxes, and Risk Queues

Sep 28, 2026

Jev labels a record that already exists. Code moves, flags, or holds it. How the document, mail, and transfer tools keep the action out of the model.

Jev Tools & CLIs#classification#email#security

Read article

Back to all resources

From the Jev community

Posts from builders shipping with Jev right now.

Follow @typesafeai