We pick and plainly summarize new features, pricing, usage limits, and policy changes across major AI tools — Claude, ChatGPT·Codex, Gemini, and Cursor — from a solo developer and maker’s point of view.
A GitHub repository now collects official MCP servers, agent skills, and agent toolkits for DevOps, cloud, platform engineering, SRE, security, IaC, observability, and diagramming work. The goal is to help people find trusted tools instead of sorting through random MCP lists. The current list covers AWS, Azure, Google Cloud, GitHub, GitLab, Azure DevOps, Atlassian, Terraform, Pulumi, Grafana, Datadog, Sentry, Splunk, PagerDuty, SonarQube, Okta, Databricks, Kubeflow, Docker, Kubernetes, draw.io, and agent-related toolkits. The list focuses on official or vendor-backed tools where possible, with notes about risk, write access, human approval, and real operations use cases. Specialized DevOps and SRE agents and reference workflows are planned next.
SkinSync is a web game that uses a phone’s front camera to read face movement. The challenges include moving your head to avoid tomatoes, opening your mouth to collect cookies, and using your eyebrows to control a small bird. Face tracking runs on the device with MediaPipe Face Landmarker, and the needed files are included locally, so the game does not call a CDN while running. Camera frames do not leave the phone, and the demo video shows only face-tracking points instead of the real camera image. The technical setup is intentionally simple: static vanilla JavaScript, no build step, and one Node file for the server. Claude Code wrote most of the code, while the human work was to play every challenge with a real face and report what felt wrong. Before release, each game also had an automated test pilot with human-like reaction delay, run without a visible screen in Node. The eyebrow-controlled flying game turned out to be impossible in its first form because about 350 milliseconds of eyebrow detection delay meant the test pilot usually passed only one gate no matter how the physics were changed, so the control was rebuilt as continuous movement where raised eyebrows make the character climb. A dependency-free PNG script made for the favicon accidentally produced a smiling face made of tracking points, and that became the mascot.
Long Claude Code sessions can use much more of a Fable allowance when they are paused for more than an hour and then continued. Claude needs to read the earlier conversation when handling a new request, so a long chat carries a large context. Prompt caching keeps that context on the server and lets later requests read it at about 10% of the normal input price. When the cache expires or is invalidated, the next request has to write the whole context again, which costs much more. For Fable 5, normal input is $10 per million tokens, a cache read is $1 per million tokens, a cache write is $12.50 per million tokens in the 5-minute window, and $20 per million tokens in the 1-hour cache window. In a heavy session with 400,000 tokens of context, a normal cached turn costs about $0.40, but returning after more than an hour can create about $8 of API cost. One expired cache can cost about as much as 20 normal cached turns.
A coding setup that has relied on Claude Code Max Pro for more than a year is now looking at local coding models. The work involves very large and complex code repositories, so the test is not just for small examples. A new Ubuntu server with a Blackwell 6000 Pro makes local model testing possible. The main question is whether there is a command-line tool that feels close to Claude Code, and which open or local coding models are worth trying first. The search is happening after about 60 days away from the space, with the expectation that model choices may have changed quickly.
Giving an AI agent a full conversation history can still lead to wrong answers. The missing piece is often not more information, but the ability to find and connect the right parts of the information. A model may find that a team moved to OAuth2, but miss the earlier security discussion that caused the change. It may also find the right sequence of events and then invent the reason behind them. A better approach is to store history as events, decisions, and clear cause-and-effect links between them. Then, when asked why the team moved to OAuth2, the AI agent can follow the chain from the security incident to the decision and rollout. The links need to be conservative: one thing happening before another does not prove it caused it, so a link should only exist when the record supports it.
A small AC-130-style game was built in one morning with Claude Code on a paid subscription. The audio and sound effects were created and adjusted on a local computer using an RTX GPU. The illustrations and character portraits were made with a local AI tool called Z-Image. The game can be played at insane.app, and it does not include Google tracking or ads. The gameplay does not follow how a real AC-130 is normally used, but the rules were changed to make the game work better. Better 3D workflows and new mission ideas are still being sought.
This MCP server turns raw app screenshots in a folder into App Store preview images. Claude checks the app’s colors, suggests visual themes and captions, then waits for approval before creating framed iPhone preview images at 1284×2778. The finished images are meant to be ready for App Store Connect. The tool is open source and can be installed with one uvx command. It was built with Claude Code, uses Pillow to draw the iPhone frame in code without separate image assets, uses a palette extractor to match the app’s colors, and exposes its features through three standard input/output tools with the official MCP SDK.
When a new high-end AI model appears, or when a company changes usage limits and safety rules, people often consider moving their subscription from one AI service to another. The harder part is not only the model itself. Each company also provides its own app, interface, features, and workflow. Those first-party tools have different strengths and weaknesses, and people build habits around them. Moving between AI services every few months can seriously disrupt an established personal workflow. The main question is whether there are setups or tools that make switching between AI platforms feel nearly seamless.
Cursor Pro Plan costs $20 per month, and extra charges can appear after the included tokens are used up. Using Sonnet 4.6 for every task can make cost control hard, especially for a beginner. Moving to Ultra Plan may help, but it does not remove the need to manage token use when Cursor is used almost every day for 4 to 6 hours. A better workflow is to split work into stages, such as planning, coding, fixing, and reviewing, then choose a suitable model for each stage. Stronger models can be saved for planning or hard problems, while simpler coding changes can use lighter models.
whatbroke is a free open-source CLI that gives Cursor better evidence when a test suite or development server fails. Pasting only a stack trace can lead Cursor to edit the wrong file because it does not know what changed since the code last worked. The tool wraps an existing command, such as `npx whatbroke run -- npm test`, and records useful failure context when the command crashes. That context includes the error, a parsed stack trace, the code changes since the last passing run, and a ranked list of likely culprit files with plain reasons. The ranking is rule-based, not chosen by an LLM. It connects to Cursor through MCP, so Cursor can ask for suspected files, edit the code, and then run the exact captured command again to check whether the fix worked, failed the same way, or failed differently. Secrets are removed before data is written to disk, and everything runs locally with no account and no telemetry. Node.js and TypeScript are the main targets, with Python and Go also supported. In a benchmark of 35 real regressions, the right file was ranked first about 90% of the time and appeared in the top three every time.
Gemini 3.5 Pro appears to have slipped again, and the delay has fed worries that Gemini is falling behind newer rival models. Google I/O pointed to a June release, but later chatter moved expectations to July 17 and then to the end of July. Some of this timing is not official and comes from screenshots and accounts with a claimed record of insider information. The broader complaint is not only the delay, but the feeling that Gemini is adding features faster than it is improving the model itself. For coding work, Antigravity still does not feel as capable as a dedicated IDE. Image generation also feels less predictable, with some prompts looping back into image creation, rejecting anything too close to popular subjects, or ending in errors. Multimodal use is another pain point: one text request with prescription box photos reportedly used 5% of a 5-hour limit. Paid users also worry that quota cuts will get worse once Gemini 3.5 Pro arrives.
Gemini 3.5 Pro was expected to arrive this week, but another signal points to a July release instead. The mood around the model is cautious. Expectations include a useful upgrade for Google Pro plan users, but not necessarily a model that beats the strongest systems from OpenAI or Anthropic. One concern is that an experimental architecture may not have produced the results Google DeepMind hoped for. At the same time, Gemini 3.5 Flash now supports native computer use. That means the model can be used for tasks that involve operating a computer interface, such as seeing a screen, clicking, and typing.
In one firsthand work example, a junior developer made a messy CSV parsing utility and it was approved in five minutes. The function was too long and the naming could have been better, but the reviewer already understood the work because the developer had discussed the vendor format at lunch and struggled with type errors during standup. Later the same day, Claude Code produced a similar-sized utility for another feature. That version looked cleaner, had better structure, and included error handling the junior code did not have. It still took more than 20 minutes to review, including checks with CodeRabbit, Bugbot, and Claude, two reads of the diff, and manual tests for edge cases that normally would not have been tested. The hard part was that the AI code arrived fully formed, with no visible history of what it tried, rejected, or misunderstood. The reviewer could not tell whether the extra caution was sensible risk control or bias against AI-generated code.
A Codex 5.6 Ultra session had 23% usage left, but both the 5-hour limit and the 7-day limit suddenly returned to 100% around the expected 3:20 AM reset time. Other people reported the same reset, so it looked less like a one-account display bug and more like a broader usage reset. One community reply said Tibo had mentioned two resets within 24 hours, but this item alone does not confirm whether that was an official policy change.
A Gemini chat can start with a prepared list of short commands, so the same long prompt does not need to be typed again each time. /HUMAN asks Gemini to make writing sound more natural and less artificial. /EL10 asks for a simple explanation that a child could follow, while /DEEPER asks Gemini to reason step by step before answering. /NOYES asks it to challenge weak points instead of agreeing too quickly. /GIVE3 asks for three clearly different options, and /TABLE turns messy information into a comparison table. /TIGHTEN makes the previous answer shorter and sharper without dropping the main points. /FLOOD asks for 20 ideas, including unusual ones. /STEPS turns a task into a numbered checklist, and /REDPEN fixes grammar, clarity, and awkward wording in one pass.
At work, 50 insurance forms had to be completed, and each form needed different information. Claude was asked to handle the job even though it was not clear it could manage that many separate forms correctly. Claude filled out all 50 forms. Every form was then checked by a person, and the completed forms were accurate.
TensorSharp is an open-source inference engine for running AI language models on a local computer. It targets Unsloth GGUF models, including examples such as Gemma4, DiffusionGemma, and Qwen3.6. It supports multimodal use cases with images, vision, and audio, plus reasoning and function tool calls. It can run on Windows, macOS, and Linux, and is designed to use the GPU when available. Its API is compatible with OpenAI and Ollama, which could make it easier to connect to existing tools. Its performance is claimed to be close to llama.cpp. It is not just a C# wrapper around llama.cpp; the inference engine is implemented from the ground up. The CPU backend runs in pure C#, and it also includes CUDA, MLX, and GGML backends, with ideas borrowed from vLLM, oMLX, and llama.cpp for caching, batching, GGUF quantization, and speed improvements.
Claude’s research function is being questioned because of the kinds of sources it may include. If weak or unreliable sources appear in the results, the whole answer becomes harder to trust. The core issue is source filtering: an AI tool that gathers information from the web needs to judge source quality carefully. The specific source at issue is not clear from the available item details.
In a firsthand case, a plan made in Claude Code was placed into Cursor, and Cursor Grok 4.5 was selected as the agent. Cursor then started Sonnet and Opus 4.8 subagents during the task. The account was on Cursor’s $60 plan, with usage freshly reset that day, and this single run used 20% of the API usage allowance. The usual workflow had been to plan with Opus and build with Composer 2.5, without noticing this kind of API usage hit before. The open question is whether this is expected behavior or a Cursor bug.
Cursor’s iPhone app was used during a child’s swimming lesson to create and merge 3 PRs across personal side projects. The concrete point is that small development tasks no longer had to wait for a laptop session. Short idle moments could become real progress on unfinished side projects. This is a firsthand experience, so it is useful as a signal, but it does not show the exact work done or how carefully the code was checked.
Claude Code 2.1.202 added a setting for the usual size of dynamic workflows, with small, medium, and large options for how many worker agents Claude should generally create. The setting is guidance, not a hard limit. Workflow-created agents now include `workflow.run_id` and `workflow.name` in telemetry, so a full workflow run can be pieced together later from recorded activity. Most of the fixes are about preventing real work from getting interrupted or lost. Claude Code fixed crashes in history search, renamed background sessions reverting after restart, Remote Control commands failing, images or files from mobile or web Remote Control being dropped, wrapped SSH login links becoming hard to click, and background agent chats entering a crash-and-restart loop. The following 2.1.203 and 2.1.204 updates continued the same theme with login-expiry warnings, a clearer manual permission mode badge, fixes for slow macOS background session switching, and a headless-session hook issue that could leave remote workers idle long enough to be stopped. Separate user reports also show that specific model choices, such as Claude Opus 4.6 with 1M context, may stop working even when the coding tool itself is otherwise improving.
Gemini was asked how many times Spain finished in the top four at the World Cup, but it returned internal working details instead of the answer. The exposed text included logic for building card-style answer screens. It showed real-looking component names such as Bento, BentoCard, and chameleon. It also included a checklist for choosing what UI to show and entity IDs from Google’s Knowledge Graph. This specific UI schema does not appear to be publicly documented.
A set of personal Mac desktop apps built with Claude fully replaced QuickBooks for one person’s finance workflow. The apps handle personal finance, accounting, tax filings, and KPI tracking. All data syncs in real time, so information entered in one app moves automatically into the other apps. This removes the need to type the same details again in multiple places. When “create invoice” is selected, the app pulls the needed information and updates everything else right away. QuickBooks felt too rigid for this workflow, while the custom tools match the way the work is actually done.
In a firsthand contracting case, the work was to pull needed information from websites. Doing it by hand would have taken a lot of time, but Claude helped create a script to handle the task. The script used almost no Claude tokens, keeping the cost around 15 cents. After giving Claude the instruction to start, the worker later received the finished results and submitted them at the end of the day. The job was simple information extraction with some field expertise, not a complex reasoning or creative task.
A user shares hands-on experience running four AI coding subscriptions side by side: Cursor at $20/month, Claude Code at $20/month, Codex at $20/month, and Opencode at $10/month. Cursor's usage resets monthly rather than on a rolling window, so there's no sudden 5-hour or weekly cutoff, and its in-house models have caught up to top-tier performance, making it the strongest all-around pick. Its Composer 2.5 model delivers 400 million tokens per month for $20, well-suited for quick edits and implementation work, and can be paired with the newer Grok 4.5 model for planning. The $20 plan is recommended for most users. For Claude Code, the Fable 5 model was still available at the time of writing (until July 12), and it excels at planning, reading existing code, and fixing awkward UI issues. Its limits skew toward a smaller 5-hour window, but that appears to translate into a larger overall weekly allowance — a frequent point of comparison against Codex, which is known for confusing limits. Claude Code also performs well for design and research work and collaborative sessions, though it hard-stops mid-task once limits are hit.
dwn.bridge is a small open-source Windows app that connects Gemini’s free web chat to a local computer. It does not use API calls; it controls the browser session with Playwright and WebView2. The setup lets Gemini read and write local files, run CLI commands, and receive command output back in the chat so it can fix errors and try again. In a direct test, Gemini was asked to build a Space Invaders-style game from scratch with no hand-written code added. The app wrote the code, compiled it, opened the executable file, found compiler errors, and kept revising. After 8 rounds, the game had animated 8x8 pixel aliens, PCM audio made in memory, color-matched particle explosions for different alien types, and bunkers that broke down pixel by pixel. The project is available under the MIT license and is still in early active development.
A longtime maker of algorave music is using AI tools to move past limits that had slowed the work for years. The work began with Ruby, then moved to Python, and later to Rust, but bigger ideas often took so much time that they stretched over years or led to long breaks from burnout. Now ideas can be typed into a phone from work, bed, or the kitchen, with remote control used to keep building and improving quickly. The main change is not making a novelty output, but finally testing music ideas that used to feel out of reach. Family reaction has been negative because the work involves AI, but the experience feels deeply useful and personal. The bad sides of AI are still real, yet this is a clear positive case for this maker.
The choice between Gemini and ChatGPT for free users is not settled by one clear winner. Many replies favor ChatGPT for reasoning, writing, image work, speed, and more useful answers. Gemini is described as stronger when it is used inside Google services such as Search, Workspace, Android, Photos, Drive, and Gmail. Gemini’s free plan is seen as generous, and some people value access to its Pro model, but there are complaints about bugs, error messages, and sessions falling back to Flash. Some people feel Gemini has become worse recently, while others are frustrated that ChatGPT can be too cautious and may avoid answering. The practical split is simple: ChatGPT looks better for hard thinking and problem solving, while Gemini may be better when the task depends on Google apps and assistant-style actions.
Claude’s models are useful, and the subscription tier makes them affordable for indie developers. The problem is predictability: if access or features appear and then disappear, serious software becomes risky to build on top of Claude. A solo builder may need to plan for sudden changes instead of treating Anthropic as stable infrastructure. The argument is that Anthropic should act more like a platform company with a clear long-term commitment. AI safety still matters, but predictable service is also necessary for developers who depend on the tool. This uncertainty makes open-weight models look more attractive once they are good enough.
Csikszentmihalyi’s idea of flow says life improves when people organize their attention so they can enter deep focus more often. In everyday AI-assisted coding, output can rise a lot, and the workday can still feel demanding, while deep focus happens less often. Writing code by hand created a tight rhythm: type, think, adjust, and stay inside the work for hours. Working with AI changes that rhythm into assigning a task, waiting, checking the result, and assigning another task. Creating something can support flow, but checking and judging AI output often does not. The practical question is whether daily AI coding can still include flow, and what that new kind of flow looks like.