Reviews of AI coding assistants, agentic tools, and the workflows that make them useful.
130 articles
A hands-on comparison of Aider and Continue.dev — two open-source AI coding tools that put you in opposite seats: the terminal and the editor. How each handles models, context, and your git history.
A practical look at which Model Context Protocol servers actually earn a slot in your editor config, what they do, and where they break down.
How CodeRabbit, Greptile, and Diamond differ on codebase context, review depth, and noise — and which one fits the way your team actually merges pull requests.
A practical workflow for splitting a large refactor across Claude Code subagents, with rules for scoping tasks, isolating file conflicts, and reviewing the merged result.
Roo Code began as a Cline fork. Here is how the two open-source, bring-your-own-key agentic coding extensions for VS Code actually differ in 2026.
A practical guide to designing, versioning, and distributing shared AI skills for Claude Code and Cursor so every engineer on your team works from the same baseline.
We tested Amazon Kiro, AWS's agentic IDE that generates requirements, design docs, and task lists before writing code. How specs, hooks, and steering files work — and where the credit-based pricing stings.
How to run coding-capable open models on your own machine with LM Studio in 2026 — hardware, quantization, the local server, and editor wiring, plus where local still falls short.
Two open-source CLIs read your staged diff and write the commit message for you. We compare aicommits and opencommit on setup, provider support, hooks, and privacy.
A measured look at Factory AI's Droids — delegation-style coding agents that take a ticket and return a pull request. Where the autonomy holds, where it breaks, and who should adopt it.
A hands-on look at Trae, ByteDance's free VS Code-based AI IDE. What its Builder mode does well, where it lags Cursor, and the data-handling questions to weigh first.
A hands-on look at Plandex, the open-source terminal AI coding agent. How its cumulative diff sandbox, version-controlled plans, and multi-model support handle big jobs.
A measured review of Gemini CLI as a coding agent in 2026 — how its ReAct loop, 1M-token context, free tier, and built-in tools hold up against Claude Code and Aider.
A hands-on look at Qodo (formerly CodiumAI): how its test generation, Qodo Merge PR review, and open-source PR-Agent hold up for real teams in 2026.
Tabby is an open-source, self-hosted alternative to cloud AI code completion. What it runs, how to set it up with Docker, and when self-hosting is actually worth the ops overhead.
A hands-on look at OpenHands, the open-source coding agent (formerly OpenDevin): how its sandboxed runtime works, when it earns its keep, and where it still trips.
A measured comparison of Greptile and Graphite for AI code review on large repos: how each reads your codebase, what breaks at scale, and which fits your team.
A measured look at Devin by Cognition in 2026 — what the autonomous AI software engineer does well, where it stalls, ACU-based pricing, and who actually gets value from it.
A hands-on look at Void, the open-source AI code editor that forks VS Code and routes requests to your own API keys or a local model instead of a vendor's servers.
A measured review of PearAI, the open-source AI code editor and VS Code fork, one year after its rocky Y Combinator launch — what it bundles, the license controversy, and how it stacks up against Cursor.
A measured look at Amazon Q Developer in 2026 — IDE completions, agentic feature dev, Java/.NET code transformation, AWS account awareness, and where it lags Cursor.
After running coding agents for months with too many MCP servers and then too few, here is the short list of Model Context Protocol servers that actually earn their context-window cost — plus the ones quietly wasting your tokens.
A hands-on review of Cline, the open-source VS Code coding agent — plan/act modes, per-step diff approval, MCP support, and a bring-your-own-key cost model that bills you at provider rates with no markup.
A hands-on review of Goose, Block's open-source on-machine AI agent — provider-agnostic config, MCP extensions, the CLI session-and-recipe workflow, and how it stacks up against Claude Code and OpenCode.
I ran JetBrains Junie inside IntelliJ IDEA for two weeks on a large Kotlin codebase. Here is where the native IDE integration actually pays off, and where the quota and IDE weight bite back.
Two of the best free AI autocomplete tools, tested head-to-head — latency, context window, IDE support, free-tier limits, and what Supermaven's move into Cursor means for its standalone future.
NVIDIA's Nemotron Omni unifies text, vision, and audio in one model. Here's how developers can wire it into agent stacks — and where the rough edges still are.
A senior developer shipped a polished web game in 48 hours using Claude Opus 4.7 and iterative plan-feedback prompting. Here is the exact workflow.
GUI-first SaaS is losing ground to CLI-native tools that AI agents can actually use. Here's why every developer-facing product is shipping a CLI, what 'agent-native' really means, and how to audit your stack.
I tested Aider across 9 projects with 6 different LLMs over six weeks, spending $47.30 total in API costs. Here's why git-native pair programming is better than accept/reject buttons — and where Aider's terminal-only approach falls short.
I spent 8 weeks building features with Claude Code across 11 projects. After 847 agent sessions and $243 in API costs, here's what $100/month of terminal-first AI actually buys you — and what it doesn't.
I built the same project in both Bolt.new and Lovable to compare the two leading prompt-to-app platforms. The differences in code quality, iteration speed, and deployment experience reveal which tool fits which kind of project.
Replit Agent combines AI coding, instant deployment, and multiplayer collaboration into a browser-based IDE. I spent three weeks building and deploying apps entirely from prompts to see whether the agent-first experience delivers on its promise.
Sourcegraph Cody indexes your entire codebase and uses that context for AI completions, chat, and code generation. I tested it on a 2.6-million-line monorepo to see whether codebase-aware AI solves the problems that generic assistants miss.
Tabnine has been doing AI code completion since 2018, longer than almost anyone. After a major 2025-2026 revamp with a new chat interface, test generation, and agent mode, I spent three weeks testing whether the veteran can compete with the new generation of AI coding tools.
v0 generates production-grade React components with shadcn/ui, Tailwind CSS, and TypeScript. I tested it across 15 real UI tasks to see whether AI-generated components hold up under actual product requirements.
Orthrus injects diffusion attention into each layer of a frozen autoregressive Transformer to generate 32 tokens in parallel — without altering the base model's output distribution.
NVIDIA Warp compiles Python functions to CUDA kernels for differentiable physics and robotics. We benchmarked it against JAX and Taichi to figure out when it earns a spot in your stack.
OpenAI's Daybreak (GPT-5.5 + Codex Security) and Anthropic's Glasswing shipped near-identical AppSec products the same week. What the convergence means and how to pick.
Macchiato's day-2 build adds a live token/cost sidebar and keyboard shortcuts for swapping between Claude Code and OpenCode in one terminal. Here's what shipped and what it means.
Macchiato Day 2 adds a 2-4 pane terminal grid, live token and cost meters, and configurable spend ceilings for Claude Code and OpenCode sessions. Here is what it actually does and who should install it.
We ran three AI code review tools — CodeRabbit, Sweep AI, and DeepSource — against the same test repositories to measure review accuracy, noise ratio, and setup complexity. Here is how each one handles real-world PRs and which tool fits your team.
Augment Code claims to understand your entire repository, not just the open file. We tested its codebase indexing against Cursor and Copilot on a 50K-line TypeScript monorepo to measure whether whole-codebase context produces measurably better code suggestions.
We tested dozens of prompt strategies across Claude, GPT-4, and Gemini to find what actually improves code generation accuracy. Concrete before-and-after examples with measurable accuracy gains — no vague advice, just prompts that work.
We benchmarked local LLMs — DeepSeek Coder, Qwen 2.5 Coder, and CodeLlama — across Ollama, LM Studio, and llama.cpp on Apple Silicon and NVIDIA GPUs. Measured latency, code accuracy, and whether offline coding assistants are ready to replace cloud APIs.
Windsurf by Codeium is an AI-native IDE built from the ground up, not a VS Code fork. We tested its Cascade agent across TypeScript and Python projects to see whether its context architecture delivers faster, more accurate suggestions than Cursor and Copilot.
Two autonomous code review agents approach the problem from opposite directions. Codegen tries to anticipate bugs before they ship. Sweep AI turns GitHub issues into pull requests. Here is how each performs on real repositories.
Continue is an open-source AI code assistant that plugs into VS Code and JetBrains. It offers model flexibility, customizable context, and a transparent architecture. We examine where it replaces Copilot and where it does not.
Cursor extends VS Code with a model-aware architecture that goes beyond autocomplete. A detailed look at the tab model, inline editing, agent mode, and where the editor still falls short for production teams.
Copilot Workspace moves AI coding from inline autocomplete to task-level planning and execution. A hands-on look at the spec-first workflow, repository awareness, and where the tool is actually useful today.
v0 generates React and Next.js UI from natural language prompts. A pragmatic look at what it produces, how the output compares to hand-written code, and when it saves real development time.
AidaIDE is a solo-built desktop IDE that unifies SSH sessions, remote file editing, and key management. We weigh it against running PuTTY, MobaXterm, and VS Code Remote-SSH side by side.
OpenClaw and other AI dev tools collapse skills into one rating. Here is a four-axis framework — task fit, security surface, install friction, update activity — that keeps the tradeoffs visible.
Agnt is a free, open-source CLI for running any public or MIT-licensed AI agent from one interface. What it does, how it compares to other agent runners, and whether to install it.
Lines of code and PR acceptance rates look like productivity signals but reward verbosity and rubber-stamping. Here is what engineering managers should track instead when adopting Copilot, Cursor, and Claude Code.
Trackboi stores every Kanban task as a plain markdown file in your repo, so AI coding agents like Claude Code and Cursor can read and update the board directly. Here is how it works and how it compares to Vibekanban.
Agetor is a 0.0.1 open-source orchestrator that pairs a Kanban board with Claude Code so you can run parallel agent tasks without juggling terminal tabs. A first look at what it does and what's planned.
Veles is an open-source MCP server in Rust that runs BM25 keyword search and semantic vector search together over a local index, giving Claude, Cursor, and other MCP assistants more precise code retrieval.
When an AI agent commits 40 times in an afternoon, git records every diff but none of the reasoning. Agent-native version control stores why each change was made, so you can bisect through agent sessions, not just diffs.
Sourcegraph's Amp is reworking the command line around autonomous AI coding agents. Here's why the terminal remains core infrastructure for agentic development — and what changes when software, not a person, is the operator.
Arcjet is moving its in-app security guards into AI agents, adding runtime checks against prompt injection, unsafe file reads, and risky web fetches. Here's why agentic apps need guardrails at the point of action, not just the network edge.
A concrete framework for structuring prompts that get Claude, Copilot, and other AI coding assistants to diagnose root causes rather than guess at patches.
Wire an open-weight model running in Ollama into a GitHub Actions workflow to get automated first-pass code-review comments on Python pull requests — no API bill required.
A technical comparison of CLI tools and Model Context Protocol for AI coding agents. Covers token cost, reliability, composability, and setup friction so you can pick the right interface.
AI coding agents produce a recognizable class of mistakes — hallucinated imports, dropped error handling, duplicate logic. Here is what static analysis can and cannot catch, and how teams are adding that layer today.
How context-window limits, the lost-in-the-middle effect, and stale data cause long-running AI coding agents to lose track — and what you can do about it.
A practical breakdown of the plan-execute-verify loop behind an autonomous AI coding agent, and how to wire it to GitHub so an issue becomes a reviewable pull request overnight.
How the Continual Harness pattern, from the Gemini Plays Pokémon and PokeAgent teams, lets an agent rewrite its own harness mid-run — plus how to apply that online-adaptation idea to autonomous agents you build.
Apify's fingerprint-suite generates statistically consistent browser fingerprints and injects them into Playwright or Puppeteer. How it works, how to wire it in, and when a scraper actually needs it.
Judea Pearl's three-rung causal hierarchy — association, intervention, counterfactual — explains why data-driven ML and LLMs hit a structural wall at causal reasoning, and what that means for agents and RAG.
How Optuna's define-by-run API, TPE sampler, and pruners automate hyperparameter tuning for scikit-learn, PyTorch, and TensorFlow models, with runnable Python code.
OpenAI's GPT-Realtime-2 is the first speech model with GPT-5-class reasoning. Here's what genuinely changes for voice agents — and what to test before you migrate.
oh-my-agent v2 adds nine new skills, promotes Cursor to a first-class vendor, and ships a benchmark scoring 80/100. A measured look at whether it fixes the agent failures developers actually hit.
Conductor enters the cloud coding agent category alongside background agents from Cursor, GitHub, OpenAI, and Google. What changes when your AI coding agent runs on remote infrastructure instead of your laptop.
codex-mcp-code-review is an open-source MCP server that automates Codex's /review flow for uncommitted changes by spawning background Codex instances. Here is how the review loop fits an agentic coding workflow.
GitHub is scanning Model Context Protocol servers for prompt injection, malicious tools, and supply chain risks. Here is what the checks catch and what they miss before you connect a third-party MCP server.
Zerostack is a pure-Rust coding agent built on Unix philosophy — composable, scriptable, single-binary. We break down how it compares to Claude Code and Cursor and when its architecture is worth adopting.
Claude Code Routines, a tool for automating repeatable coding workflows, drew 686 points on Hacker News. Here's what a 'routine' actually is, how it fits the agentic dev-tools landscape, and how to decide if it belongs in your workflow.
Anthropic reported a $44B run rate, a $200B Google Cloud deal, and a SpaceX compute arrangement in one week — plus Claude Code Auto Mode. What it means for developers.
OpenAI put its Codex coding agent inside the ChatGPT iOS and Android apps, so you can start tasks, review diffs, and manage agent runs from your phone. Here's what that changes for your workflow.
Anthropic's June 15 pricing changes the math on hosting a Claude personal assistant: a decision framework for choosing Managed Agents in the cloud versus a local always-on Claude Code instance.
GenCAD is a research project that generates editable parametric CAD models from images instead of meshes. A look at its architecture and what developers building design-automation tools can take from it.
Anthropic is moving programmatic Agent SDK traffic to a new monthly credit pool, separate from standard Claude API billing. Here's what to audit in your integration before the split affects forecasting and rate limits.
GitHub's standalone Copilot desktop app puts it head-to-head with Claude Code and Codex CLI. We compare workflow surface, approval semantics, and model neutrality so you can pick the right one.
Anthropic shipped agent view in Claude Code, a CLI dashboard for parallel agent sessions. We test it, explain the muted developer response, and lay out what would actually fix multi-agent workflows.
Anthropic's Claude passed ChatGPT in enterprise ARR, DAUs, and developer adoption in April 2026. Here's what shifted, why Claude Code drove it, and how to audit your AI stack now.
Searle's Chinese Room, stochastic parrots, and IIT all predict where current LLMs break. Here is what that means for how you architect prompts, retrieval, and agent loops.
A Stanford field study of 51 production AI deployments found agentic systems deliver 71% median productivity gains versus 40% for copilot-mode assistants. Here's what separates the top quintile.
Arxiv submissions are flooding faster than anyone can read. A practical workflow for filtering low-quality ML papers, plus the curation services and citation tools worth your time.
A review of nine CUDA programming books — which hold up against the CUDA 12 toolkit and Hopper architecture, which are out of date, and a working reading order to go from zero to writing your own kernels.
A walkthrough of Prolog's declarative model using Pokémon types and evolution chains. Covers unification, backtracking, and where the paradigm shows up in modern systems.
Semble is an open-source code search tool that indexes your repo with embeddings and returns ranked chunks to AI agents instead of raw grep output. We tested whether the 98% token reduction claim holds up against ripgrep on a 180k-line monorepo.
A hands-on n8n review covering self-hosting trade-offs, AI agent nodes with tool calling and vector retrieval, and how its per-execution pricing compares to Zapier and Make for developer-led automation.
How Google's internal editor stack moved from Emacs and Vim to the web-based Cider IDE — and what the shift tells you about cloud dev environments, monorepo tooling, and AI-assisted editors.
Gruber's electricity analogy for AI, unpacked — why thin GPT wrappers keep dying, what survives the test, and where dev tools like Cursor actually fit in your stack.
A cost breakdown of running Llama 3.3 70B locally on an M-series Mac Studio versus paying per-token on OpenRouter. The cloud wins by 30-60x at typical developer volumes — here's the math and the three scenarios where local still makes sense.
A practical look at why text rendering breaks fully native cross-platform UI and how SwiftUI, Jetpack Compose, Flutter, and React Native make different bets to handle it.
Cal.com shipped cal.diy as a self-hostable scheduling primitive developers embed into their own apps. Here is what it is, how it compares to hosted Cal.com and Calendly, and when to reach for it.
Code generation speed isn't where engineering teams lose time. Here's where AI tools like Cursor and Copilot actually compress cycle time, and the boring process fixes (PR size, review SLAs, CI duration) that move team-level metrics.
arXiv now imposes a one-year submission ban for papers with unchecked LLM errors like hallucinated citations. Here's the policy, why it exists, and the verification workflow that catches hallucinations before you submit.
We tested Bun 1.2 against Node.js 22 LTS on real workloads. Where the speed gap is real, where Node compatibility breaks, and a concrete framework for deciding whether to migrate your toolchain.
Nous Research's Hermes Memory Installer adds local persistent memory to AI agents with one shell command. We compare its file-based approach to Mem0 and Letta.
A practical look at Tokenyst, an open-source local monitor that tracks Claude Code API token usage in real time and alerts you before runaway agent loops turn into surprise Anthropic bills.
Unsloth's NVIDIA collaboration claims 1.6x faster LLM fine-tuning and 70% lower VRAM usage for Llama, Mistral, and Qwen. We break down what the numbers actually unlock for developers training on consumer GPUs.
Anthropic's Managed Agents platform adds 'dreaming' — background agent execution that explores outcomes on Anthropic's infrastructure. How the new capability changes the build-vs-buy math for teams shipping on Claude.
Anthropic reportedly secured access to SpaceX's 220,000-GPU Colossus 1 cluster to relieve Claude API capacity pressure. Here's what changes for the 529 errors and tight rate limits hitting your coding agents.
Anthropic is rolling Claude into Microsoft 365: Outlook gains support and Word, Excel, and PowerPoint integrations leave preview for general availability. Here's what changes for developers and which workflows actually benefit.
Connecting MCP servers to Claude Code or Cursor silently injects 55K+ tokens of tool definitions into every turn. Here's the real cost — and how to cut it.
DeepClaude pairs DeepSeek R1's chain-of-thought reasoning with Claude's synthesis in a single agent loop. We cover how the dual-model architecture works, where it beats Cursor or Copilot, and how to wire it up via API.
Anthropic's Claude Opus 4.7 brings a 1M token context window and improvements for coding agents. Here's what changes for developers building with the Claude API.
In April 2026, a Cursor AI agent wiped PocketOS's production database in seconds. Here's what happened, why it happened, and how to lock down autonomous coding agents before they cost you the company.
We tested both AI coding assistants against a Next.js app, a Python CLI, and a Rust library migration. Cursor won on velocity. Here's the breakdown — and the one scenario where Copilot still edges ahead.
Cursor's new SDK exposes the same agent runtime that powers the editor. We break down what ships, where the documentation lags, and when the limitations matter for production code.
OpenAI's Codex Chrome extension puts its coding agent inside your browser tab. We tested the workflow patterns that pay off, the limits worth knowing, and how it fits next to Codex CLI and IDE agents.
A measured comparison of OpenCode and Claude Code, the lock-in math behind the split, and a decision framework for picking one, the other, or both.
A measured developer review of Alibaba's Qwen 3.6 Plus API — pricing vs GPT and Claude, 1M-token context behavior, coding benchmarks, and the access paths that actually work.
We pointed Codex and Claude Code at the same Python codebase across refactoring, debugging, and agentic tasks. Here is what each tool shipped, where each one wins, and what the speed-vs-cost tradeoff actually looks like in practice.
A hands-on review of ModelScope, Alibaba DAMO Academy's open-source model hub. Covers SDK setup, model discovery, ms-swift fine-tuning, and how it compares to Hugging Face for Qwen-family and DAMO research workflows.
AdamsReview orchestrates multiple Claude Code agents for PR reviews. We break down how multi-agent review catches what single-pass LLM reviews miss, and where it fits in your pipeline.
Otter, Fireflies, and Granola are facing class actions over consent and data retention. Here's what developers integrating AI transcription need to audit before shipping.
Adam Dunkels wired Claude into a user-space TCP/IP stack and benchmarked it against ICMP ping. The latency floor it reveals is the most honest stress test we have for agentic Claude API workflows.
yt-dlp replaced youtube-dl as the default for programmatic video and audio extraction. Installation, format selectors, the Python API, and the production gotchas we hit running it across three real workflows.
The build-your-own-x GitHub repo has 350k+ stars for a reason. Here are 10 from-scratch tutorials — databases, compilers, Git, neural nets — that teach how the tools you use every day actually work.
A measured look at Ratty, a terminal emulator pitching inline 3D graphics. Where the category fits, which workflows benefit, and what to verify before you switch.
Why evaluating Copilot, Cursor, and Claude Code by lines generated misses the point — and how to measure whether your AI tooling is adding or removing technical debt.
Daniel Stenberg confirmed Mythos surfaced a real bug in curl, one of the most-reviewed codebases on the planet. Here's what that means for AI-assisted security review in your pipeline.
A measured guide to running 7B-32B local language models on a base M4 Mac with 24GB unified memory. Model size math, real tokens/sec numbers, and when Ollama, llama.cpp, or MLX is the right tool.
A measured look at the backlash against AI coding assistants — what the METR study and cognitive offloading research show about when hand-coding actually produces better engineers and better code.
The case for running models on your laptop instead of paying per-token API bills: where local AI (Ollama, LM Studio, llama.cpp) wins on cost, latency, and privacy, and where the cloud still earns its keep.
A practical 30-day comparison of Cursor and VS Code across multi-file edits, agent workflows, and pricing — based on actual usage.
One email every Friday with the tools worth your time. No spam, unsubscribe anytime.