- unwind ai
- Posts
- Bring ChatGPT and Claude-like Features to Any LLM
Bring ChatGPT and Claude-like Features to Any LLM
+ Advanced tool use in Claude, 30 pro tips for using Gemini CLI
Shubham Saboo & Gargi Gupta
November 28, 2025
In partnership with
Today’s top AI Highlights:
& so much more!
Read time: 3 mins
AI Tutorial
Imagine uploading a photo of your outdated kitchen and instantly getting a photorealistic rendering of what it could look like after renovation, complete with budget breakdowns, timelines, and contractor recommendations. That's exactly what we're building today.
In this tutorial, you'll create a sophisticated multi-agent home renovation planner using Google's Agent Development Kit (ADK) and Gemini 2.5 Flash Image (aka Nano Banana).
It analyzes photos of your current space, understands your style preferences from inspiration images, and generates stunning visualizations of your renovated room while keeping your budget in mind.
We share hands-on tutorials like this every week, designed to help you stay ahead in the world of AI. If you're serious about leveling up your AI skills and staying ahead of the curve, subscribe now and be the first to access our latest tutorials.
Latest Developments
Building a proper AI assistant means juggling RAG systems, web search APIs, agent frameworks, and trying to make them all work together nicely.
Onyx bundles everything into one self-hostable chat UI - MCP integration, knowledge graph RAG, web search, deep research, custom agents, code execution, and connectors to 40+ sources, working seamlessly with any LLM from proprietary to self-hosted.
The interface handles OpenAI, Anthropic, Gemini, or local models through Ollama and vLLM with the same smooth experience, no code changes needed when switching providers. Deploy with a single curl command or go full enterprise with Kubernetes, SSO, and RBAC, and run it completely airgapped if your industry demands it. Teams are using it to serve thousands of employees with department-specific assistants, each using the best model for their workflow while pulling from the same knowledge base.
Key Highlights:
One Platform, Every Feature - Get RAG with knowledge graphs, web search via Google PSE or Exa, MCP integration, code execution, image generation, and custom agents without stitching together separate tools.
40+ Source Connectors - Pull knowledge and metadata from Google Drive, Slack, Notion, Confluence, and dozens more, with built-in indexing that keeps retrieval fast at scale.
Deployment Options - Run via Docker, Kubernetes, Terraform, or cloud-specific guides for AWS EKS and Azure VMs, with options for complete airgap deployments using locally hosted LLMs.
Model Agnostic - Switch between OpenAI, Anthropic, Gemini, or self-hosted options like Ollama and vLLM, without rebuilding your setup, all through the same interface with consistent tooling.
Tech moves fast, but you're still playing catch-up?
That's exactly why 100K+ engineers working at Google, Meta, and Apple read The Code twice a week.
Here's what you get:
Curated tech news that shapes your career - Filtered from thousands of sources so you know what's coming 6 months early.
Practical resources you can use immediately - Real tutorials and tools that solve actual engineering problems.
Research papers and insights decoded - We break down complex tech so you understand what matters.
All delivered twice a week in just 2 short emails.
Your AI agent just tried to load 58 tools and burned through 55,000 tokens before writing a single line of code.
Anthropic's new advanced tool use features fix this mess by letting Claude discover tools on-demand, orchestrate multi-step workflows through Python instead of repeated API calls, and learn correct usage patterns from examples rather than guessing from schema definitions.
The three beta features - Tool Search Tool, Programmatic Tool Calling, and Tool Use Examples work together to solve the core bottlenecks that make production agent development painful: context bloat from tool definitions, token waste from intermediate results, and parameter errors from ambiguous schemas.
Internal testing shows these aren't just incremental improvements. Opus 4.5 jumped from 79.5% to 88.1% accuracy on MCP evaluations, while complex research tasks saw 37% token reduction and knowledge retrieval improved from 25.6% to 28.5%.
Key Highlights:
Tool Search Tool - Reduces token overhead by 85% through on-demand discovery, loading only 3-5 relevant tools (~3K tokens) instead of entire libraries upfront (~72K tokens), while regex and BM25 search capabilities help Claude find the right tools from thousands available.
Programmatic Tool Calling - Lets Claude write Python orchestration code that processes tool outputs inside a sandbox, keeping intermediate results out of context. A budget compliance check that would traditionally load 2,000+ expense line items now returns just the 2-3 people who exceeded limits.
Tool Use Examples - Provides concrete usage patterns directly in tool definitions, teaching Claude format conventions (YYYY-MM-DD dates, USR-XXXXX IDs), nested structure handling, and parameter correlations that JSON schemas can't express. Accuracy improved from 72% to 90% on complex parameter handling.
Parallel Execution & Latency Gains - Eliminates 19+ inference passes when Claude orchestrates 20+ tool calls in a single code block, with async support letting scripts fetch data from multiple sources simultaneously instead of sequential API round-trips.
Quick Bites
30 Pro tips for effectively using Gemini CLI for agentic coding
If Gemini CLI has been sitting in your terminal unused, this guide by Google’s engineer, Addy Osmani, walks through the features that make it stick. His guide includes ~ 30 tips, like GEMINI.md files that persist project context so you stop repeating the same instructions, /memory commands that accumulate useful facts across sessions, and MCP server integrations, that’ll supercharge your workflows with Gemini 3 Pro in the Gemini CLI. Think of it as the setup guide that turns a free-tier terminal agent into something you'd actually reach for daily.
Microsoft open-sourced a 7B agentic CUA model
Microsoft just dropped Fara-7B, a 7B parameter agentic model that can autonomously use your computer, clicking, typing, and navigating to complete tasks. It operates visually by perceiving UIs and taking appropriate actions. Built on Qwen2.5-VL and trained on 145K synthetic trajectories from Magentic-One, it's punching above its weight class: outperforming UI-TARS and even competing with OpenAI's computer-use-preview. MIT-licensed and available on Azure Foundry and Hugging Face, plus optimized for on-device deployment on Copilot+ PCs.
FLUX.2 matches Nano Banana Pro and runs locally
Black Forest Labs just dropped FLUX.2, their latest image gen family that can generate high-quality images and handle up to 10 reference images simultaneously. The open-weight FLUX.2 [dev] model combines text-to-image and multi-reference editing in a single 32B parameter checkpoint, and their pro and flex variants offer production-ready endpoints with variable step controls for balancing quality and speed. While Google's Nano Banana Pro is getting buzz for its reasoning-backed infographics and 4K outputs, FLUX.2's architecture leverages Mistral-3's 24B VLM for world knowledge plus a rectified flow transformer for spatial logic. You can actually run the dev version locally on a single RTX 4090.
Andrey Ng’s agentic paper reviewer matches human reviewers
Andrew Ng just released an agentic paper reviewer that's hitting human-level performance - 0.42 correlation with human reviewers versus the 0.41 correlation between humans themselves. The system takes your research paper, searches arXiv for relevant prior work, and generates a grounded review with actionable feedback in minutes instead of the typical 6-month peer review wait. Try it at paperreview.ai, though it works best in AI fields where research lives openly on arXiv.
Build with Jules without a GitHub repo
Google’s async coding agent Jules now works without a GitHub repo attached, which means you can finally stop creating throwaway repositories just to test an idea. Upload whatever context you need code snippets, files, screenshots, and let the agent build in a clean sandbox before you download the results. It's the "napkin sketch" mode that async coding agents should have shipped with from day one.
Tools of the Trade
Jarvis - Open-source and local version of Wispr Flow. A powerful, privacy-first voice dictation and AI assistant for macOS. Press a hotkey, speak, and your words are transcribed and pasted instantly. Your API keys stay on your machine. Uses Deepgram Nova-3 or OpenAI Whisper STT models.
AI Workflow Builder - Vercel's open-source template for building workflow automation platforms with a drag-and-drop interface that outputs production-ready TypeScript code marked with "use workflow". Ships with pre-built integrations for Resend, Linear, and Slack, and runs on Workflow DevKit with full deployment support.
Momentic - AI-powered end-to-end testing platform that lets engineers write tests in plain English rather than code. It uses intent-based locators that survive DOM changes. It also comes with an autonomous agent that can explore your app, generate tests, and handle non-deterministic outputs like LLM responses.
Awesome LLM Apps - A curated collection of LLM apps with RAG, AI Agents, multi-agent teams, MCP, voice agents, and more. The apps use models from OpenAI, Anthropic, Google, and open-source models like DeepSeek, Qwen, and Llama that you can run locally on your computer.
(Now accepting GitHub sponsorships)
Hot Takes
Startup Idea:
An IDE that shows you ads while your agent is spinning
and the ads pay for the tokens
~ Humiyeah anthropic was right, coding is probably done for pretty soon
opus 4.5 + composer 1 + gemini 3 is basically an engineer at this point (backend/design/ iteration)
That’s all for today! See you tomorrow with more such AI-filled content.
Don’t forget to share this newsletter on your social channels and tag Unwind AI to support us!
PS: We curate this AI newsletter every day for FREE, your support is what keeps us going. If you find value in what you read, share it with at least one, two (or 20) of your friends 😉
Reply