A Vibe Coding assistant for VS Code that goes far beyond simple chat. Mendes AI turns your editor into a specialist across four full areas:
- General coding — explain, refactor, generate, fix code in any language
- Data analytics — inventory datasets, frame business questions, drill into root causes, generate decks
- Data Science — clean, EDA, build/train/tune ML models, explain them with SHAP/LIME
- AI Engineering — production RAG, prompt engineering, LoRA fine-tuning, LLM evaluation, agents
All powered by large open-weight models (Qwen, Gemma, custom GGUFs) running on Google Colab Pro GPUs, streamed to your VS Code over a Cloudflared tunnel. No local GPU required.
Architecture in one diagram
┌─────────────────────┐ ┌──────────────────────────┐
│ VS Code (you) │ │ Google Colab Pro (L4/A100)│
│ ┌───────────────┐ │ HTTPS/SSE │ ┌────────────────────┐ │
│ │ Mendes AI │◄─┼───tunnel──────┼─▶│ FastAPI router │ │
│ │ sidebar chat │ │ │ │ + Ollama runtime │ │
│ └───────────────┘ │ │ └────────────────────┘ │
│ - 14 slash skills │ │ Qwen 27B · Gemma 26B │
│ - Tool calling │ │ (or your own GGUF) │
│ - Voice + vision │ │ │
└─────────────────────┘ └──────────────────────────┘
The big models live where there's VRAM. Your editor stays light. Tools execute only on your machine — the backend just emits intents.
What you can do — the four areas
🧑💻 General coding (works everywhere)
- Open Chat: dedicated sidebar chat with a Mendes AI icon in the Activity Bar (same UX as Copilot Chat / Claude Code). Click and use, no setup ritual.
- Inline commands:
Mendes: Explain Selection(Ctrl+Alt+E)Mendes: Refactor Selection(Ctrl+Alt+R)Mendes: Generate at Cursor
- Tool calling: the model can read/write files, run bash, grep, glob, fetch the web, etc. Every tool call shows a confirmation modal with Accept / Always / Deny. Dangerous bash patterns (
rm -rf /,dd, fork bombs) block the "Always" option as a safety net. - Voice input and image attachments for vision-capable models.
- Persistent chat history across sessions.
📊 Data analytics — /explore, /question, /chart, /drill, /report
Adapted from ai-analyst-lab/ai-analyst.
| Slash | What it does |
|---|---|
/explore |
Inventory a dataset, profile every column, surface quality blockers (nulls, duplicates, outliers), recommend feasible analyses |
/question |
Turn a vague business problem into 5-10 testable analytical questions ranked by Impact × Feasibility, with hypotheses for the top 3 |
/chart |
Generate Storytelling-with-Data styled charts — action titles, max 2 colors + gray, direct labels, no clutter |
/drill |
Root cause investigation with "peel the onion" methodology: confirm → decompose → isolate → repeat — until you find the specific actionable cause |
/report |
Consolidate the full analysis into a business-ready final_report.md + executable_analysis.ipynb |
🔬 Data Science (ML) — /engineer, /eda, /model, /explain
Adapted from HungHsunHan/claude-code-data-science-team.
| Slash | What it does |
|---|---|
/engineer |
Data cleaning, missing values, outlier detection (IQR / Z-score), duplicates, feature engineering driven by business value |
/eda |
Exploratory analysis — descriptive stats, distributions, correlations (Pearson/Spearman), pattern discovery, hypothesis formulation |
/model |
Model selection (linear / trees / boosting / NN), train/val/test splits, hyperparameter tuning, evaluation metrics matched to the business goal |
/explain |
Multi-method feature importance: built-in + permutation + SHAP + LIME + Partial Dependence + interactions + stability checks |
🤖 AI Engineering — /rag, /prompt, /finetune, /eval, /agent
Original to Mendes AI — built from current best practices.
| Slash | What it does |
|---|---|
/rag |
Production RAG pipelines: chunking strategies, embedding selection, hybrid retrieval (semantic + BM25), reranking, prompt template, evaluation with Ragas |
/prompt |
Production prompt engineering — Chain-of-Thought, few-shot, structured outputs (JSON/XML), function calling schemas, anti-jailbreak, robustness testing |
/finetune |
LoRA / QLoRA fine-tuning end-to-end: dataset prep, Unsloth/transformers setup, hyperparams, eval against base, deployment as merged model or adapter, GGUF quantization |
/eval |
LLM evaluation systems: reference-based + rubric-based + reference-free, LLM-as-judge done right, Ragas / DeepEval / PromptFoo / LangSmith, CI integration |
/agent |
Agent architecture — tool catalog design, ReAct / plan-and-execute / reflexion, memory tiers, error handling, framework picking (LangGraph / LlamaIndex / AutoGen / CrewAI / Smolagents) |
Type /skills or /help in the chat to list all 14 with descriptions.
Quick start
- Install — search Mendes AI in the Extensions panel (
Ctrl+Shift+X) or use:code --install-extension mendesalex89.mendes-ai - Spin up the backend — open the Colab notebook and run all cells. The last cell prints:
MENDES_URL=https://xxxxx.trycloudflare.com MENDES_TOKEN=... - Configure VS Code — open Settings → search "Mendes AI":
mendes.tunnelUrl→ the Cloudflared URLmendes.token→ the token
- Click the Mendes AI icon in the Activity Bar (the brain + chip in the left rail) — the chat opens in the sidebar.
- Type
/skillsto see everything you can do.
Settings
| Setting | Default | Purpose |
|---|---|---|
mendes.tunnelUrl |
"" |
Cloudflared tunnel URL from your Colab session |
mendes.token |
"" |
Bearer token sent on every request (machine-scoped, doesn't sync) |
mendes.toolsEnabled |
false |
Allow the model to call tools (file I/O, bash, web) |
mendes.healthPollSeconds |
10 |
How often to poll /health for the status bar |
Keybindings
| Action | Shortcut |
|---|---|
| Focus Mendes chat (sidebar) | Ctrl+Shift+M (Cmd+Shift+M) |
| Explain selection | Ctrl+Alt+E |
| Refactor selection | Ctrl+Alt+R |
Privacy & security
- Tools require explicit per-session approval. "Always" decisions live in RAM and reset when VS Code closes.
- Token is machine-scoped — does not sync across devices via Settings Sync.
- All traffic flows through your own Cloudflared tunnel and your own Colab session. No third-party server in the middle.
- File system access and shell execution happen only on your machine — the backend just forwards the model's intent.
Why this stack?
- VRAM economics. Qwen 27B and Gemma 26B don't fit on an 8GB local GPU. They fit on Colab Pro's L4 / A100.
- Zero per-token cost. Colab Pro is a flat monthly fee — vs. metered API billing.
- Privacy. Prompts never leave a Colab kernel you control.
- Editor stays local. All file I/O and shell execution happen on your machine.
Limitations
- Needs a Colab Pro account (or another GPU host) to run the backend. The extension does not include hosted inference.
- The Cloudflared tunnel URL changes when Colab restarts — you must update the setting (the extension shows a clear error when it's offline).
- AI Engineering skills generate code; they don't run training/RAG/eval pipelines for you. They tell you what to write and why.
Issues & contributing
Bug reports, feature requests, and PRs are welcome at github.com/mendesalex89/AI_Assistant.
Acknowledgments
Skills are adapted from these excellent open-source projects:
- Data analytics skills from ai-analyst-lab/ai-analyst (Apache 2.0)
- Data Science skills from HungHsunHan/claude-code-data-science-team
License
MIT — see LICENSE.