Welcome back. AI is moving from our screens to our plates. A UK-based company just dropped an ingredient-embedding model that learns the flavor geometry from over 4 million recipes across 7 languages. It maps out how nearly 1,800 ingredients interact, and the makers claim the whole flavor map fits in roughly 2MB, smaller than a phone photo.

Also: How to fine-tune your own LLM, find out the update on Codex's next default model, and see why OpenAI just bet $250M on AI and jobs.

OpenAI connects private servers without exposing them: The ChatGPT maker just shipped Secure MCP Tunnel, a way to connect private MCP servers to its products without opening firewall ports or exposing data. You run a lightweight client inside your network that makes an outbound HTTPS connection to handle requests. That finally lets teams with strict security rules use ChatGPT, Codex, and the Responses API with their internal tools.

Anthropic updates Claude Code for speed and stability: The AI lab just pushed a batch of reliability fixes to its terminal coding agent. A new full-screen renderer stops the flickering that used to ruin long sessions. Thinking and tool calls now stream in real time, so Claude no longer looks frozen mid-task. Compaction now shows progress to head off prompt-length errors. MCP connections are more stable, and crashed sessions can recover on their own.

Ramp deploys 10,000 agents to scan its codebase for vulnerabilities: The fintech company ran ten thousand parallel coding agents across its backend and fixed several high-severity issues from the results. The scan pipeline is model-agnostic and runs on publicly available models. Ramp benchmarked GPT-5.5, Kimi K2.6, and DeepSeek V4 Pro against its confirmed vulnerabilities, with GPT-5.5 leading at 52% detection, but the open-weight options still catching real issues. Any team running background coding agents can adopt the same playbook.

Mailtrap’s OpenClaw integration adds an Email Sandbox layer to your agent’s workflow.

Route outgoing email to a sandbox first, where you can safely inspect and validate agent behavior before it ever reaches a real user.

Install the Mailtrap skill, add your API token and Inbox ID, and test the integration to get started!

Start sending AI email safely

DeepSeek breaks the AI playbook. DeepSeek is the Chinese AI lab that sent shockwaves through the market last year, triggering a $1T sell-off in US tech stocks with its affordable, open-source models. Since then, they've ignored Silicon Valley’s playbook by giving away their architectural secrets and skipping the usual subscription models for coding, video, and audio. It has investors wondering how they actually plan to turn a profit.

The “Big Fund” is leading the round. The Chinese AI lab is in talks for a $10B funding round at a $45B valuation. Interestingly, the round is reportedly led by China’s state semiconductor fund: a group that usually backs chipmakers.

Architecture is the strategy. The biggest bottleneck in AI is specialized memory, mostly controlled by South Korean and US suppliers. DeepSeek's models slash those memory requirements, letting older Chinese chips run cutting-edge AI. Alibaba and Tencent ordered hundreds of thousands of next-gen Huawei chips before DeepSeek's V4 launched. DeepSeek gave Huawei early access to optimize V4 for inference and shut Nvidia out of that, but the model was still trained on Nvidia hardware, so the dependency it's trying to escape is only half-broken.

Open-source weights are a long-term play. By offering free model weights, DeepSeek is driving massive demand for Chinese suppliers. While they walked away from subscription revenue, the goal is to capture a piece of a domestic AI hardware market that could be worth trillions. If you want to dive deep into the technical details behind the move, this original viral thread (1M+ views) breaks down every architectural choice and the strategic logic behind it.

Start building with:

Use code THECODE30 for 30% off your first purchase.

Get API Key

Code review feedback is usually asynchronous, forcing you to break your flow just to address comments and wait for the next round. One engineer from OpenAI’s Codex team shared a way to automate this. By using Codex’s new thread automations, an agent can poll your PR and draft fixes while you stay focused on your work.

To set it up:

Codex now checks for new comments on a schedule, drafts fixes in the background, and only pings you when necessary.

P.S. Get 50+ AI coding hacks for Claude Code, Cursor, and Codex here.

How to use Codex’s /goal command: This tutorial shows you how to use the “/goal” feature in Codex to run autonomous, long-running tasks. You'll discover how to write outcome-driven prompts with specific success criteria, letting the AI fix errors and wrap up complex projects on its own without you having to micromanage it.

Buildpipe: A local-first automation tool for developers who want to chain shell commands, AI calls, and file operations into reusable pipelines. You can trigger them on a schedule, a file change, or via webhook.

Claude Code Harness: This repo creates a structured delivery pipeline for Claude Code, integrating specific paths for Codex and OpenCode. It offers a repeatable, test-first workflow that bakes in specs and evidence so every update ships with confidence.

Building self-improving tax agents with Codex (OpenAI): Fixing real-world AI errors usually requires a slow, manual engineering process. By transforming expert corrections into structured feedback, Codex-driven agents can autonomously self-improve, rapidly boosting system accuracy and saving significant time.

Anthropic just unveiled a new security plugin for Claude Code that integrates vulnerability reviews directly into the development workflow.

Grow customers & revenue: Join companies like Google, IBM, and Datadog. Showcase your product to our 300K+ engineers and 150K+ followers on socials. Get in touch.

You can also reply directly to this email if you have suggestions, feedback, or questions.

Until next time — The Code team