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.
Cursor can show that a codebase is 100% indexed while the agent run still shows no clear use of semantic search. The visible logs showed grep searches and file reads, not semantic search, vector search, or RAG. When asked directly, the Opus 4.8 and Composer 2.5 agents first said they did not have a semantic search tool and listed only the tools they had. The codebase had about 250 files, so size did not clearly explain the behavior. Later, semantic search seemed to start working suddenly, and the agents then said they had direct access to the semantic search tool.
Gemini says its 1 million token context window can understand up to about 1,500 pages of text or 30,000 lines of code. In a firsthand test with a 400-page book PDF, both Flash and Pro produced only one-sentence summaries for chapters later in the book when asked for a detailed chapter-by-chapter summary. Pro also showed a warning that the uploads might be too large for best results. Even on the Ultra plan, uploading a very long file does not guarantee deep, even coverage of the whole document. The 1 million token claim appears to describe how much Gemini can take in, not a promise that every long-document task will produce complete, high-quality output.
Cursor failed to return answers from AI agents while running in Auto mode. The error said the model name was not valid. The model name shown in the error was "ex-tahoma-riverheart-xhigh". This points to a possible problem in Cursor’s automatic model selection, rather than a normal prompt or coding mistake by the user.
In a firsthand account, access to Cyberuse was removed after switching from a paid OpenAI plan to a free plan. The email notice said access would return after moving back to a paid plan. The paid plan was restored, but Cyberuse verification still failed. The error message said to contact support if it was a mistake, but support replied that it could not fix the issue. The result was a paid account with the feature still blocked and no clear next step.
Cursor’s cloud agents helped save time and money during a few weeks of real development work. Bugbot was useful for finding bugs in the code that got built. But it did not check whether the finished work matched the original plan. Missing requested features and extra work outside the intended scope still needed separate review. The workflow shifted from reviewing only the code changes to reviewing the plan first. Then a second agent compared the finished build against that plan and flagged anything that had drifted. Internal tooling was built around this review process.
AI coding tools such as Claude Code, Codex, and Cursor often lose useful working memory when a new chat starts or when the context is compressed. In one Claude assistant setup named Igor, the practical pain was that the assistant felt new again after about 200,000 tokens, and reloading Telegram chat logs around 160,000 tokens still did not create lasting memory. The wider pattern is moving from simple prompt compression toward separate memory and context layers that preserve project structure, past decisions, rules, and recent work between sessions. Some tools try to save tokens by loading only the tools that matter for the current task. Others aim to stop agents from rediscovering the same repository details every time they start. Aictx reports avoiding about 4,000 to 13,000 tokens of repeated repository rediscovery per prompt, while another MCP server claims a 46% reduction in AI coding agent token costs. A related track focuses on control and audit: tools such as Aegisure try to apply one set of rules across multiple agents and check local code changes for risks like secrets, payment or login changes, and skipped tests before code is pushed.
Claude and GPT-5.5 are given the same coding, debugging, or analysis prompt at the same time, then a fresh Claude combines the two answers without knowing which model wrote which one. The aim is to avoid choosing only one model and instead produce a stronger final answer from both outputs. GPT-5.5 runs through Codex CLI using an existing ChatGPT subscription, so the workflow avoids a separate API key and metered token bill. Inside Claude Code, the process is packaged as a single /fuse command. A blind benchmark found wins in some coding, debugging, and analysis cases, but the method did not win every time. The workflow is framed as an intermediate repeatable process for quality control, token saving, context and memory handling, debugging, and multi-agent work.
Agency Agents App is a manager for the Agency Agents repository, which has more than 113,000 stars. It lets people browse more than 230 AI agents by division and role, then inspect each persona before installing it. The app can deploy the same kind of agent setup into Claude Code, Codex, Cursor, Gemini CLI, Qwen, opencode, and Copilot. It also tracks what it wrote, detects when settings drift, and lets users update or roll changes back. The older setup used a shell command and menu system, while this app gives a more controlled visual interface. It is open source under the MIT license, with no telemetry, accounts, or signups.
git-lrc is a free tool that runs a small AI code review when code is saved as a Git commit. The problem it targets is simple: heavier use of AI coding tools can produce more code while developers spend less time reading and understanding what was generated. That can let regressions reach production and later require a rollback. git-lrc moves review earlier than the PR stage, while the change is still fresh in the developer’s mind. On commit, it opens a review screen with the diff, summarizes what changed, points out areas that deserve another look, and lets the developer jump through important parts of the change. It now checks about 100 common risk patterns across 10 areas, including security, reliability, performance, and maintainability.
Claude can be connected to a bank through MCP so business money tasks are handled by conversation. The described workflow includes invoicing, bill pay, expense tracking, and bookkeeping. The main idea is to ask Claude to do the work instead of moving through banking and accounting screens by hand. Comments made it sound like the setup works reasonably well, but daily use, exact setup steps, and reliability are still unclear.
VibeAround is an open-source development tool for running several AI coding agents side by side on one computer. It supports Claude Code, Codex, Gemini CLI, OpenCode, desktop agents, provider profiles, terminals, browser previews, and remote follow-ups. The tool is meant to solve a common messy setup: separate terminal windows for each agent, different command-line tools, separate settings files for model providers, local previews spread across different ports, and disconnected use across browser, terminal, editor, and phone. Before starting work, it lets someone choose the workspace, saved session, terminal, agent, and API profile. It also lets people switch provider and model profiles without repeatedly editing global settings. It manages local browser previews, Markdown and HTML previews, and web terminal sessions. Agent work stays local by default, and a desktop coding session can be continued from a phone or messaging channel.
ChatGPT used to separate memory use into two choices. One choice used only saved memory, and another used both saved memory and recent chat history. The setting to turn off chat history reference is no longer visible in this case. The visible option is now “try improved memory.” Even when that option is turned off, ChatGPT still appears to use recent chat history, which is not the desired behavior. The preferred setup is to use saved memory only and ignore recent chats.
When building with Cursor, there is a need for one command or workflow that lays out the next build steps. Optional improvements that are pushed to later can disappear from both the developer's memory and Cursor's working context. The goal is to keep one main roadmap, epics, and todo list so future work does not get scattered across chats or memory.
An experiment is underway to build a Skyrim-style role-playing game using AI tools only. No code has been written by hand, and the visible result has been created through prompts. The first few hours went into a long starting prompt that set the game world, tone, and visual style. After that, the work moved into repeated changes to details such as animations, collision handling, combat, and items. About 20 hours have gone into the build so far, and it is still far from finished. The goal is a large role-playing game with quests and dragons, and the current version already has a look and feel the maker is happy with.
HSBC estimates that OpenAI must grow annual revenue from about $13 billion to $200 billion by 2030 and raise another $200 billion to stay financially stable. In the same broader picture, OpenAI’s losses in 2025 reportedly grew almost eightfold, while spending reached $34 billion. The main issue is that ChatGPT and developer AI tools are growing fast, but the computing power and infrastructure needed to run them are extremely expensive. OpenAI needs many more paying users, business customers, developer usage, and outside funding to keep scaling. These numbers suggest that popularity alone is not enough; prices, plan limits, enterprise sales, and long-term funding will matter more over time.
Cursor Mobile is presented as coming soon, and the main question is how useful an AI coding tool can be on a phone. The strongest use case is not writing a whole app on a small screen. It is sending small fixes, planning tasks, continuing a coding conversation from a laptop, or asking an agent running at home to start work before returning to a computer. Some people see this as useful for quick ideas, iPad work, and on-the-go checks. Others argue that real coding needs a large screen and that changing a business website from a phone is too risky. There are also concerns about burning through a token budget quickly. The discussion also raises voice input, iOS and Android support, and comparisons with Codex mobile and Claude Code.
ChatGPT voice mode may be close to an upgrade. A new model name, gpt-bidi-1, has been spotted, along with update text that appears connected to voice features. The exact changes are not confirmed yet. People are asking how it would differ from gpt-realtime-2, and some are hoping voice mode becomes smarter and more useful. There is no confirmed release date or clear feature list yet.
Cursor showed what appears to be a new model during COMPILE26. The screen said the model has more than 1.5 trillion parameters and was trained from the ground up on more than 100,000 GPUs. The slide also framed it as useful beyond coding, not just as a code assistant. People in the Cursor community guessed it could be Composer 3, while others debated whether it was related to Grok, Kimi, or DeepSeek. Some pushed back on that idea because the slide said the model was trained from scratch. Reactions were mixed: some were excited, some wanted proof that it is better than Composer 2.5, and some worried that a more expensive model should not replace the current cheaper choices.
manim-video is a coding-agent skill for making math and technical explainer animations from simple prompts. A person describes the concept, then the coding agent follows a workflow for planning, creating scenes, rendering, stitching scenes together, and reviewing the result. Docker handles the runtime, while Manim creates the animation. This means the user does not need to install Manim, LaTeX, or ffmpeg directly on their computer. The skill is meant to work with tools such as Codex, Claude Code, Antigravity, OpenCode, and other coding tools that support agent skills. One example use case is animating a proof that the angles in a triangle add up to 180 degrees.
In a firsthand case, Claude Opus 4.8 Max refused to draft an answer in a format that could be placed directly into a work document. The request appears to have been a normal writing task, but Claude seemed to reject it because the text was not related to AI. The exact refusal text and full chat context were not provided. This is a check for whether others have seen the same behavior, not yet proof of a broad bug or policy change.
Developers are pictured as moving between AI coding tools like Claude and Codex at the start of each month. Usage quotas run out, a new model launch creates excitement, and slightly better benchmark scores make another tool look more attractive. Work spent tuning a CLAUDE.md file for Claude is abandoned when developers move to Codex. A 50% discount appears during subscription cancellation, showing how AI tool companies try to keep users from leaving. At the same time, another group is moving back from Codex to Claude, which makes the main point clear: many developers keep cycling between tools instead of settling on one permanent choice.
Otto is an open-source tool that lets Claude control a real Chrome browser tab. It includes an MCP server, so Claude can call browser actions such as opening a tab, moving to a page, extracting information, taking a screenshot, and intercepting network requests. It works with a live Chrome tab instead of using a headless browser setup or renting a cloud browser. The main design goal is token efficiency. Fixed code performs the browser actions, while Claude only decides what the next step should be. The tool is released under the MIT license and can be installed globally with npm. Its creator has been using it to read and extract information from sites that block headless browsers.
Fujifilm camera recipes are often shared with beautiful sample photos, but they can look very different when applied to someone else’s own photos. A small app was built with Claude in about 45 minutes to solve that problem. The app lets a person upload a photo, apply camera recipe changes, and see the result live. The screen puts the photo preview at the top and the controls underneath in a scrollable area. It also includes a button to change the photo, a way to save recipes, a recipes view, and an export option for a recipe card. The maker had no coding knowledge and created it through a back-and-forth prompt conversation with Claude for a React app.
MEX is an open source tool for giving AI coding tools smaller, more targeted project context. It creates a `.mex/` folder in the project root. Instead of one large context file, the agent starts from a short bootstrap of about 120 tokens that points to a routing table. The routing table chooses the right context file based on the task. For example, an auth task can load the architecture file, while new code work can load the conventions file. This means the agent reads the information it needs and avoids extra material. MEX also includes drift detection for checking whether the scaffold still matches the real codebase. It uses 8 checkers without AI or token use, and can catch missing file paths, deleted npm scripts mentioned in docs, dependency version conflicts across files, and scaffold files that have not changed for more than 50 commits. When it finds problems, `mex sync` builds a focused prompt for Claude Code to fix only the broken files.
The Finance Toolkit developer has added an MCP Server on top of the project. Finance Toolkit is an open-source Python package started in 2019, with more than 200 financial metrics, valuation models, and economic indicators. The project began because major finance sites showed different P/E numbers for the same company on the same day, often without explaining the calculation method. The new MCP Server groups more than 200 methods into about 21 tool categories, so Claude, Copilot, Cursor, Windsurf, or Gemini can fetch real financial data and run calculations instead of relying on remembered training data. Setup uses one command: `uvx --from "financetoolkit[mcp]" financetoolkit-mcp-setup`. Full documentation and an MCPB file for Claude Desktop are available. In one example, a request to compare major semiconductor companies over 10 years led the AI tool to choose valuation ratios, EPS growth, and historical return tools on its own.
The personal site nicolae.tech is being rebuilt with the Payload Website Template. Payload handles the main site pieces, including pages, hero sections, layout blocks, media, posts, SEO, redirects, and forms. The goal is not a fixed brochure site, but a CMS-driven site that can grow over time. The structure separates the hero area from normal page content and uses reusable blocks that still need to feel designed rather than generic. Real credits and case studies are meant to build trust quickly. The site should later expand into blog posts, case studies, and EventBlok content. Claude Design and Claude Code are being used alongside existing Payload knowledge to make the template more personal and useful. The open questions are whether the site still looks too much like a template, whether the page structure fits Payload well, whether the CMS model should change, and what should improve before more case studies are added.
Non-technical people can run into friction when using Claude Code for everyday work. They may need to pick the same folder again and again, find projects confusing, and struggle with memory that does not clearly update or come back when needed. This desktop app treats projects as the main starting point. A user creates a project, links a folder once, and works from there. Memory is meant to update quietly in the background, and the app shows a visual graph of stored memories. The built-in agent is named Rick and has a chosen personality, with room to add other agents with different personalities. Task management is included inside the app, so users and agents can track work without switching to Notion. The app is built with the Claude SDK and Claude 4.8, with a beta planned for waitlist users in July.
A small macOS menu-bar app detects when Claude Desktop or Claude Code has stopped because of a usage limit. It reads the reset time shown by Claude, waits until that time, then automatically types “continue” so the conversation can resume. It can track several paused conversations at the same time, with a separate timer for each one. The app keeps a local activity log, can start when the Mac starts, and stays out of the Dock. It is described as having no network access and no telemetry. The code is available under the MIT license, and the installer is offered as a DMG on GitHub.
Kery is an autonomous QA tool for web apps. A developer points it at a URL, runs `npx keryai`, and it crawls routes, forms, and modals to map what is actually in the app. Tests are written in plain English, such as checking whether checkout finishes successfully, and Kery decides how to carry them out. It is built to avoid the upkeep that often comes with Playwright scripts, CSS selectors, and XPath. Instead of relying on brittle page-code locations, it uses the accessibility tree and screenshots to understand what buttons, fields, and screens mean. That should make tests less likely to fail just because a developer moved or renamed something. In a nightly staging workflow, Kery reports broken flows, visual problems, and user-experience issues in a dashboard with screenshots and marked problem areas. A triage agent uses memory from earlier runs to reduce duplicate bug reports and false alarms.
Relaticle, an open-source CRM, added an in-app AI agent in version 3.3. The agent can read CRM data and propose actions such as creating companies, updating deals, and attaching notes. It does not change data on its own; every write action needs human approval first. The stack uses laravel/ai for the agent layer, Reverb for live streaming, Filament v5 and Livewire v4 for the interface, and Horizon for background jobs. The hardest part was making live chat responses survive normal product problems. The chat runs as a queued job and streams through Reverb, but users may reload the page, lose the websocket connection, or trigger a Livewire re-render. Each stream needs its own identity, and the app must be able to resume a half-finished answer after reconnecting. In production, route caching also caused broadcast channel authorization to stop registering silently, which took time to find.