"CLI coding agents: OpenAI Codex, Claude Code, Gemini CLI"

I just published a short blog post about Socraites, a Laravel package I’ve been building to explore how AI can help with code reviews. It covers how I approached building AI context, the challenges I faced, and what I learned along the way. https://lnkd.in/d6CDmnf8

I’ve become a big fan of the OpenAI SDK — not because it’s flashy, but because it’s well engineered. In a world full of abstraction-heavy GenAI frameworks, this one feels refreshingly honest. Sharing a few thoughts on why it deserves more love 👇 The OpenAI SDK: Simplicity Done Right If you haven’t tried the OpenAI SDK, it’s worth a look. In the vast ocean of GenAI frameworks and agent builders, this one doesn’t get nearly the love it deserves. Everyone’s talking about complex orchestration layers, agent graphs, and “frameworks for frameworks” — but sometimes, all you need is a clean, mature SDK that just works. And that’s exactly what this feels like. You can tell it was built by engineers who actually write production code. It has that rare quality: it feels designed by experienced developers, for developers. 💡 A Breath of Fresh Air I’ve tried a lot of GenAI frameworks lately. Many are impressive, but also… heavy. They hide simple ideas behind layers of abstraction — “chains,” “contexts,” “graphs,” “memory stores.” Great for demos, not so great when something breaks at 2 a.m. The OpenAI SDK is different. It’s light, clean, and honest about what it does. No unnecessary magic. No guessing. Just solid engineering. ⚙️ Why It Stands Out Mature API design — clear, predictable, and consistent. Simplicity — minimal boilerplate, maximum clarity. Easy to debug — what you see is what’s happening. Performant — no hidden overhead or invisible queues. Fits anywhere — drop it into an ETL job, FastAPI route, or notebook and it feels native. It’s a rare SDK that scales from quick prototypes to production without forcing a rewrite. 👩💻 Built for Developers There’s confidence in its restraint. It doesn’t try to own your stack. It just gives you the right primitives to build your own architecture. That’s the hallmark of thoughtful software design. 🚀 Growing — Without Growing Pains Even as OpenAI adds new capabilities — like the Agent SDK and Agent Builder — the philosophy stays consistent. Each layer adds power without adding clutter. You don’t have to learn a new paradigm; it just fits naturally into what you already know. That’s rare — and it shows real design maturity. ❤️ It Deserves More Love Maybe because it’s not flashy, the OpenAI SDK often flies under the radar. But it’s elegant. It’s dependable. And it’s built with respect for developers who care about clarity and control. If you’re building serious GenAI systems — not just demos — it’s absolutely worth a closer look. 🔜 Coming Up Next In my next post, I’ll dive deeper into the Agent SDK and Agent Builder, and explore how OpenAI is redefining what “agentic” can mean — without the bloat. Sometimes the best tools aren’t the loudest — they’re the ones that just fit the way you think. 💬 Curious what others think — have you tried the OpenAI SDK yet? What’s your take compared to other GenAI frameworks?

openai just did an ama on reddit and i read through the entire thing here's some the most interesting points that stood out > just shipped gpt-5-codex model specifically optimized for agentic coding with focused training on diverse coding tasks >model performs significantly better than base gpt-5 in coding environments due to specialized optimization >team dogfoods heavily - "i use codex to write 99% of my changes to codex" with goal of zero manual coding next year team usage & productivity impact > engineers can prototype large features with just ~5 turns of prompting, building 3 different feature versions in single day > designers with no coding background now directly merge prs using codex, eliminating traditional developer bottleneck > massive productivity gains reported: "no more writing crud endpoints or stream helpers" - focus shifts to higher-level architecture > enables rapid experimentation where suboptimal prototypes can be scrapped without emotional attachment since they're "cheap to build" performance trade-offs & complaints > new gpt-5-codex significantly slower than previous models (15-20 minutes vs 3-4 minutes for same tasks) > more intelligent but speed decrease breaks iterative development workflows for many users technical capabilities & roadmap > cli supports web search with --search flag, coming to ide extension soon with prompt caching fixes > planning mode in development after internal debates about user control vs model autonomy > exploring browser automation and voice integration (inspired by open source community demos) > considering sub-agents and workflow automation but nothing actively in development platform & integration challenges > currently focused only on vs code due to limited resources, jetbrains users frustrated with no timeline > no free tier planned, codex remains paid-only despite requests windows experience problematic with excessive permission prompts and powershell command issues token economics & rate limits > teams working with 3M-10M token codebase mentions how gpt-5 is "magical but we are just getting started" for massive codebases pro users ($200/month) still hitting same rate limits as plus users for codex specifically > confusing limit descriptions like "300-1500 messages per 5 hours" when single session can run 25 minutes > no usage tracking ui, users blindsided by multi-day lockouts mid-workflow but they said its coming future vision & philosophy > tools should become invisible infrastructure like compilers that "fade into the background" > abstraction level rising - engineers will work at system level rather than writing individual functions > optimistic scenario: engineers own larger product spaces, need broader generalist skills across design/product/engineering

🧑🏻💻 Here are 4 zero-cost methods for AI-powered coding in VSCode using Kilo Code ( Kilo Code ) a flexible alternative to GitHub Copilot that supports custom OpenAI-compatible APIs. Method 1: Qwen-Code CLI + Free OAuth Use Alibaba's Qwen-Code CLI, a tool optimized for code understanding with the Qwen3-Coder model. You get 2,000 free API calls daily with a 60 requests/minute limit and no token caps. Setup for Qwen-Code: 1. Install via npm: npm install -g @qwen-code/qwen-code 2. Run qwen and authenticate via your browser with /auth. 3. In Kilo Code settings, set Provider to "Qwen Code" and Model to qwen3-coder or qwen3-coder-plus. Method 2: ModelScope API ModelScope, Alibaba's model community, offers 2,000 free daily inference calls for models like Qwen3-Coder after linking an Alibaba Cloud account. It's fully OpenAI compatible. Setup for ModelScope: 1. Register on modelscope.cn, link your Alibaba account, and create an API Token. 2. In Kilo Code: • Provider: OpenAI Compatible • Base URL: api.modelscope.cn/v1 • API Key: Your ModelScope Token • Model: qwen3-coder Method 3: OpenRouter Free Models OpenRouter aggregates models like DeepSeek V3 and Gemini 2.5 Flash. A one-time $10 top-up (your balance isn't used for free models) unlocks 1,000 free calls daily (20/min limit). Without it, you get 50. Setup for OpenRouter: 1. Register on OpenRouter add $10 credit, and generate an API key. 2. In Kilo Code: • Provider: OpenRouter • API Key: Your OpenRouter Key • Model: A free model like deepseek-v3:free or gemini-2.5-flash:free. Method 4: Docker Self-hosted Gemini Balance If you have multiple free Gemini API keys from Google AI Studio, you can self-host a load balancer to avoid rate limits. Gemini Balance deploys easily via Docker and is optimized for Gemini 2.5 Pro. Setup for Gemini Balance: 1. Pull the Docker image: https://lnkd.in/gWX9eP4d 2. Configure a .env file with your Gemini keys and access tokens. 3. In Kilo Code: • Provider: OpenAI Compatible • Base URL: http://<your-server-ip>:8000/v1 • API Key: Your access token By load-balancing multiple free Gemini keys, you can achieve nearly unlimited high-frequency AI coding assistance, tailored to your needs. These four methods give you powerful, truly free AI coding capabilities right inside VSCode. #VSCode #FreeAI #QwenCode #ModelScope

I've been having relatively good success using agentic AI CLI tools like Claude Code (Anthropic) and Codex CLI (OpenAI) for coding tasks. The fact that they have access to the full power of the shell, my full project context, and integrate with my preferred and familiar Jetbrain IDEs (Rider, Webstorm, etc), makes for a much better workflow than cut & paste code snippets. There is a continuity in conversation around the code base, and it does start to feel much more like a coding partner. Of course, it is critical to use good software engineering practices with these new tools just as one would with any project - i.e. always review code diffs before committing to your git repo, do proper testing, etc. Also, their hallucinations and gaslighting over-exuberance about half-baked or even completely broken code can be a grating nuisance. But even with these limitations, I do find myself able to tackle more in less time. YMMV. As I started spending more time in this mode, however, it became very jarring to abruptly hit up against API limits. My coding partner would suddenly take a multi-hours-long coffee break in the middle of a productive session! So, I had an idea - what if I used Claude Code to create a system that would let me easily continue a coding session where I left off with GPT-5 or some other LLM? Also, I was curious to test the recent open source models, and even see if there were any LLMs that would run at all well locally on my machine (TL;DR - no, an Apple M1 with 16 gigs doesn't quite cut it, but it was fun to explore) The result is this open source project, AI CLI Switchboard: https://lnkd.in/g5jkQKsP (Github: PatchedReality/ai-cli-switchboard) Currently supported CLIs: * Claude Code * Codex CLI Remote APIs: * OpenRouter * DeepSeek * Google Gemini * Z.AI GLM-4.5 (Claude Code only) Local models: Any local MLX or LM Studio model (proxied through LiteLLM) You can add connections to new models just by adding a yaml configuration file. (Just ask Claude to add it for you) For this project, I purposely was liberal with allowing the AI's to write nearly all of the code in the repo, including the git commit comments and README, to see how far it could get. I even used the tool itself while in development to switch between models to continue development on parts of it. Through the process, I learned a good deal about MLX, LiteLLM, LM Studio, the limitations in how LLM's access local tools, and the various coding LLM's that currently exist. I hope it can serve as a useful resource for others that want to kickstart their own exploration of what is possible, and find the tools that work for them in the process. If you give it a try, I'd love to hear what you think! (Btw, I wrote this post myself)

My Learning Journey with RAG (Retrieval-Augmented Generation) for Code Understanding ---> Over the past few days, I’ve been working on RAG (Retrieval-Augmented Generation) and applying it to a very practical use case: making codebases more understandable with the help of LLMs. <-- What I Built --> I implemented a pipeline where a query is asked about a codebase (in my case, a Python file), and instead of feeding the entire file directly to the LLM, I used a retrieval mechanism to surface only the most relevant code snippets. The final response is generated based strictly on that retrieved context. **Here’s a breakdown of the components I used** -- Text Splitting -> Leveraged RecursiveCharacterTextSplitter to chunk Python code logically using custom separators like class and def. -- Embedding & Retrieval -> Used sfr_embedding_code_400m_r_model from Salesforce (Open Model) for generating embeddings and retrieving top-k relevant chunks. **Prompt Engineering** - Created a custom PromptTemplate that ensures the model sticks to the given code context and avoids hallucination. **Runnables & Chain** - Built a RunnableLambda to dynamically pass retrieved code snippets into the prompt, then pipelined it with an OpenAI LLM and a StrOutputParser for clean outputs. **Final Execution** - Queried the system with questions like “Merging of File and explain the code” and received contextually accurate answers extracted from the code. <--- Key Learnings ---> **Smart Retrieval > Dumping Context ** - Feeding the entire codebase is inefficient and often exceeds context limits. Retrieval ensures only relevant pieces are used. **Custom Splitting is Crucial** - Default text splitting often breaks code in unnatural ways. Adding language-specific separators (class, def) made retrieval much more meaningful. **Hallucination Control** - Explicitly instructing the LLM not to answer if context is missing helped improve reliability. **Composable Chains in LangChain** - Using RunnableLambda and chaining made the entire flow modular and easy to extend. <--- Why It Matters ---> Firstly I am thankfull to Nitish Singh for the Guidance and creating a Spark in me to create something on Ai Domain. As codebases grow, understanding and navigating them becomes a huge challenge. A RAG-powered approach can act like a “smart coding assistant”, answering specific queries with code references, explanations, and insights—without overwhelming the LLM with irrelevant context. This experiment is just the beginning, and I’m excited to explore more advanced retrieval strategies and integrations for real-world developer workflows. Github - https://lnkd.in/gZK9BCDH Curious to know how others are approaching RAG for code. Let’s discuss! #RAG #LangChain #LLM #CodeUnderstanding #AI #LearningInPublic

Claude Bug Postmortem, and what it means to your AI agents: A few weeks ago, during my late-night coding, I was very close to smashing my keyboard on the screen. Claude is against me. That’s it. Like literally he’s telling me: “Go to sleep. You coded too much for one night. I’m tired too.” Did you also feel it? We weren’t hallucinating. Anthropic just published a postmortem (link in first comment): there were 3 separate bugs causing Claude to act weird. Here’s what really happened: 1) Requests were misrouted Some prompts went to the wrong servers. Imagine asking your smartest colleague a question, but half the time the message gets forwarded to an intern. You still get an answer, just not the one you needed. → Impact: outputs looked inconsistent, sometimes strangely off-topic, sometimes lower quality than usual. 2) Corrupted outputs A config bug injected garbage characters into responses by mistakenly assigning higher probability to rare tokens, causing broken code, random tokens, nonsense in the middle of otherwise fine text. → Impact: developers got code that wouldn’t even compile, or text that looked like Claude had a glitch mid-sentence. 3) Token-selection bug in the TPU compiler An optimization bug sometimes caused Claude to literally skip over the right answer. → Impact: you’d get a confident but totally wrong response. Each of these alone would be painful. Overlapping together, they made it feel like Claude suddenly got stubborn and dumb (or both). Now, what does it mean for your AI agents? Your agents don’t just rely on prompts and logic, they also rely on the LLM and its vendor. These third-party LLMs can have unexplainable and unnoticed bugs. It’s not like a normal software bug that crashes loudly or obviously returns garbage. This is an LLM: sometimes you don’t even know it’s wrong, you just have that unexplainable feeling that the world is against you. After testing and evaluating your agents the best you can, and putting all the required guardrails, they can start to behave differently and operate differently in production. And until you even notice, you’re screwed. When the LLM has this kind of bug (vs. any other “traditional” software bug): - The output doesn’t crash loudly. It looks fine. - Guardrails don’t save you, because nothing “unsafe” happened, just different (again our famous “allowed but wrong” thing). - Your careful pre-prod evals won’t help here. This is a runtime issue happening live. So remember: agentic behavior is the new blindspot.

Software eats the world, AI reprograms the menu. That’s the vibe as OpenAI unleashes their latest in coding intelligence, GPT-5 Codex. Picture this: a virtual co-pilot that not only writes boilerplate, but can now refactor your gnarly legacy code or architect full-stack web apps without breaking a sweat. Novice? Doesn’t matter. “Vibe coding” lets you wield intent in plain English and watch Codex translate it into code. But, there’s a twist, the AI code-gen race already has a market king: Anthropic’s Claude Code. OpenAI’s move isn’t just an upgrade; it’s a statement. GPT-5 Codex posts an impressive 74.5% on SWE-bench Verified, leaping past its predecessor’s code refactoring prowess from 33.9% to 51.3%. We’re not looking at incremental gains here; we’re seeing a redefinition of what’s possible for dev productivity. But Codex isn’t just for the pros, its Terminal, IDE, and Web flavors mean anyone can pilot it, regardless of whether you can tell Python from Perl. That’s both liberating and transformative for non-coders and seasoned architects alike. Suddenly, every user is an application builder. Industry-wide, these advances are seismic. Code generation tools are evolving from monolithic prompt engines into specialized agents. AI-assisted workflows will become the new normal. Competitors, brace yourselves. Tool vendors, rethink the value you’re offering: IDEs will need to double down on AI integrations or risk obsolescence. However, if this is the new baseline, developers must ask new questions. How do we trust generated code in production? Where’s the accountability in a world where anyone can launch an app? As the race between Claude and Codex heats up, enterprise security and code review pipelines will require robust adaptation to keep pace. The bottom line: AI code agents are no longer just helpers; they’re teammates at the keyboard. Tech leaders should start prepping for upskilled teams tasked with validating and securing auto-generated software, while non-tech business units can begin leveraging AI to prototype faster. Whether you’re a veteran coder or a business leader eyeing digital acceleration, keep close tabs on this next wave; the boundaries of who creates, maintains, and secures software just changed. More details and benchmarks in the original coverage: https://lnkd.in/g5DXfbCP (OpenAI's new GPT-5 Codex model takes on Claude Code) Don't just take our word for it, read the full story here: https://lnkd.in/g5DXfbCP #SECURITYOPERATIONS #BLUETEAM #CYBERSECURITY #SOC #DIRECTOROFAI

Your operating system runs on binaries. Your AI agents should run on Markdown. A traditional OS executes /usr/bin/python—compiled, opaque, rigid. You can't edit a binary mid-execution. Markdown-as-OS: No binaries. Just Markdown files. Plain English prompts. An AI agent reads them, executes them—and edits them while running. September 2025: The shift GitHub open-sourced Spec Kit for spec-driven development (Claude Code, Copilot, Gemini CLI). Core idea: Start with a Markdown spec, make it source of truth, let agents generate/test/validate code from it. AGENT.md standard gained native support across Amp, Claude Code, Cursor, Firebase Studio, Gemini CLI, OpenAI Codex. "The README your AI coding agents actually read." The Markdown OS structure: /specs/ → requirements (AGENT.md, feature.md) /plans/ → agent-generated execution plans /runs/ → logs (markdown + NDJSON journal) /state/ → queue.md (task checkboxes), deployment state /reports/ → human-readable summaries Workflow: Write specs/feature.md → Agent generates plans/ → Executes, logs to runs/ → Updates state/queue.md (checks tasks) → Writes reports/summary.md All Markdown. No compilation. Just English instructions agents read, execute, and edit. Task queue example: ## High Priority - [x] Deploy contact form ✓ 2025-09-30 - [ ] Fix auth bug (assigned: auth_agent) Agents grep for - [ ], claim tasks, update to - [x], commit. Why it wins: • Greppable: grep "TODO" specs/*.md • Diffable: git diff state/queue.md • Version-controlled: Full audit trail • Agent-friendly: LLMs parse Markdown natively • Human-readable: Any text editor • Self-modifying: Agent edits its own instructions mid-execution 2025 validation: LLMunix ("Pure Markdown OS"), agentic.md (Markdown + Mermaid agents), GitHub Spec Kit, Medium: "The Markdown Revolution" Comparison: Binary executable → Markdown prompt Compiled code → No code, just instructions Opaque config → AGENT.md (human-readable) State in DB → State in markdown files Rigid (recompile) → Dynamic (edit mid-run) Real workflow: Marketing writes specs/product-launch.md → Agent reads spec → Generates plan → Executes (build, test, deploy) → Logs to runs/journal.jsonl → Writes reports/summary.md Time: 30 minutes. The spec was clear. The plan was visible. The execution was logged. All in Markdown. 2025 consensus: Markdown is the universal interface between humans and agents. — Want the starter kit? DM "MARKDOWN OS" for the full repo template (AGENT.md + queue system + Python agent + example specs).