Autonomous AI-powered research loop for VS Code. ResearchLoop automates the full ML research cycle: literature review, code extraction, experiment design, hyperparameter tuning, and report generation.
- Literature Search — Searches arXiv, Semantic Scholar, and DuckDuckGo for relevant papers. Supports custom sources (any website via domain-scoped search).
- Paper Analysis — LLM-powered analysis of papers for relevance, methods, and key findings.
- Code Extraction — Extracts reusable code snippets from discovered papers and repositories.
- Experiment Engine — Iterative experiment loop with automatic hyperparameter exploration, reflection between rounds, and early stopping.
- Checkpoint & Resume — Pause and continue experiments across sessions without losing progress.
- Results Analysis — Automated statistical analysis of experiment results.
- Report Generation — Produces Markdown or LaTeX reports summarizing the full research cycle.
- Telegram Notifications — Get notified on pipeline events and control your pipeline remotely via bot commands (
/status,/pause,/continue,/restart,/new). - Skills System — Persistent user instructions injected into LLM prompts (global or per-workspace).
| Provider | Model (default) | Notes |
|---|---|---|
| Ollama | llama3.1 | Free, local, no API key needed |
| Claude | claude-sonnet-4 | Requires Anthropic API key |
| OpenAI | o3-mini | Requires OpenAI API key |
- Install the extension
- Open the ResearchLoop panel in the Activity Bar
- Click + to create a new research session
- Enter your research question and target metrics
- Click Run — the pipeline handles the rest
Open ResearchLoop: Settings from the command palette to configure:
- LLM provider and API keys
- Literature sources (built-in + custom URLs)
- arXiv categories filter
- Experiment limits (max experiments, early stopping threshold)
- Telegram bot notifications
- Skills (persistent instructions for the LLM)
- Create a bot via @BotFather on Telegram and copy the bot token
- Open a chat with your bot and send
/startto initialize the channel - Visit
https://api.telegram.org/bot<YOUR_TOKEN>/getUpdatesin your browser - Copy the
chat.idvalue from the JSON response - In VS Code settings, set:
researchloop.notifications.telegram.enabled:trueresearchloop.notifications.telegram.botToken: your bot tokenresearchloop.notifications.telegram.chatId: the chat ID from step 4
- Reload VS Code — you should receive a "ResearchLoop connected" message
Commands: /status /pause /resume /stop /continue /continue N /restart /new
If you open multiple VS Code workspaces with Telegram enabled, messages are prefixed with
[workspace-name]so you know which instance is talking. Each instance processes commands independently.
This extension is in early development (beta). Features may change, break, or behave unexpectedly. Use at your own risk — always review generated code before running experiments on sensitive data. Feedback and bug reports are welcome via GitHub Issues.
- VS Code 1.85+
- For local LLM: Ollama running on localhost
- For experiments: Python 3.8+ with scikit-learn (installed automatically if missing)