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.
Shellular is a remote control tool for using a full development setup from a phone. It runs on the user’s own Mac, PC, Mac mini, or VPS, not only inside a separate cloud workspace. It supports AI coding tools such as Claude Code, Codex, OpenCode, Cursor CLI, and Pi with a mobile interface for slash commands, file mentions, approvals, and model selection. It can keep terminals running without tmux. It also includes an in-app browser and DevTools, so localhost pages can open without setting up port forwarding. Setup is described as running `npx shellular`, scanning a QR code in the app, and approving the connection in under two minutes. Because everything runs on the user’s own machine, the AI tools keep access to the full code context. The connection is described as end-to-end encrypted, and the app is available on both the App Store and Play Store.
Agentic coding tools such as Claude Code, Cursor, and Codex have made it much easier to start building software. A person with internet access can now describe changes in normal language and get large codebase-level edits. Hobbyists, junior developers, and product managers can create much more software than before. This could lead to a huge increase in the amount of software being made. Some of it may be excellent, but a lot of it may be messy code that only works because AI kept patching it together.
SuperGrok is available for 10 CAD, while ChatGPT Plus and Gemini Pro each cost 28 CAD in this case. The main decision is which service is better for everyday use and coding or software development. The comparison depends on model access, usage limits, and how useful each plan is for real development work. No benchmark, feature list, or hands-on result is provided, so the clear facts are the prices and the intended use.
things3-cloud is a command-line tool for managing Things 3 tasks. It is written in Rust and connects directly to the Things Cloud API, so AI agents like Claude Code or Codex can read and update tasks. Its goal is to avoid local workarounds such as AppleScript, Shortcuts, URL schemes, or reading the local database. Each command syncs changes to the cloud right away. It supports common actions such as viewing today’s tasks, searching by text and deadline, and marking tasks done. It can also create and edit tasks, manage projects, areas, and tags, update checklists, search tasks, and use a local cache for faster startup. The main risk is that it relies on an unofficial API discovered by reverse engineering. Earlier incorrect API requests caused official Things clients to behave badly and made Things Cloud access appear frozen, though it could be recovered by turning Things Cloud off and on again.
Claude can be used as a repeatable work system, not just as a place to ask one-off questions. Reusable prompts can be turned into “skills” so tasks like code review, project management, debugging, and context handling become more consistent. One pattern is to split work across multiple Claude agents, have them review code from different angles, then loop through fixes and checks. Another pattern is to run checks before and after a build so requirements and quality are not missed. For long tasks, forced state management helps Claude keep track of past decisions, current progress, and next steps. Sandboxing can let Claude test, correct itself, and review its own changes in a safer space. Simpler text transformation tools and CLI-style command flows are also part of the approach.
Pi coding agent can be combined with Herdr and Gondolin to split one project across several AI agents. Herdr is used as the workspace, with the first tab acting as the orchestrator agent that directs the rest. A custom extension called Herdr-control lets that orchestrator create agents inside Gondolin micro VMs and place them into chosen Herdr workspaces, tabs, and panes. The setup can group agents by role, such as three researcher panes in a research tab, four planner panes in a planning tab, and a worker in a worker area. The micro VMs are sandboxed, but all agents share a central memory directory so they can pass information to each other. Discussion around the setup raised an important missing detail: the model and cost matter, not just the tools. One reply noted that the screenshot appears to show worker agents using Mimo 2.5.
Gemini AI credits may have a pricing problem where buying more at once can make each credit more expensive. Bulk buying usually suggests a lower price per credit, but the larger bundle may not always be the better deal here. The available information does not confirm the exact prices, purchase flow, country, or which Gemini product this applies to.
Google’s full-stack AI means offering the main parts needed to build and run AI products as one connected system. Those parts include computing infrastructure, an AI model, a platform that coordinates work, and the screens or apps people use. Google points to its TPU chips, Gemini models from Google DeepMind, Gemini Enterprise Platform, and everyday products such as Gmail and Maps as parts of this approach. The claimed benefit is that builders do not have to connect many separate vendors by themselves, and Google can manage reliability across more layers of the system. Google also says this setup is not meant to be fully closed, because other AI models or outside software can still be connected. Suggested starting points are Google AI Studio for quick web app prototypes, Gemini Enterprise Platform for low-code work automation, and Antigravity for more complex apps or agents.
A team built its own CLI because many internal apps did not provide an official one. That CLI has worked well for AI agent workflows, almost like a local MCP. The new concern is a plan to use the same CLI inside the main backend logic of web apps. The proposed setup would have a Python backend start a child process, run CLI commands, and let those commands make the API calls. The existing approach has been to call the API directly through an SDK or Python request library. The concern is that the CLI route adds harder error handling, retries, extra memory use, and more moving parts. Codex and Claude Code may make code easier to write, but that does not make an agent-friendly tool a good production backend design. The argument in favor is that the commands already work with local AI agents, so reusing them may feel easier and simpler to debug.
Claude is being considered as a main tool for several solo building tasks. The planned use includes controlling or connecting several services with MCP, coding in the terminal with Claude Code, building apps and websites, and working on multiple projects each month. The work is not described as heavy enterprise-scale use, but the goal is to avoid frequent usage limits while building. The main concern is running out of tokens or hitting rate limits often enough to interrupt the work. The practical choice is whether Pro is enough or whether Max is the better fit.
Every Claude conversation since January was saved and later analyzed with a clustering tool to see how usage had changed across work and personal life. The expected result was about 30 different uses. The actual result was 5 core themes appearing in 38 different professional situations, and 4 of those themes were unnamed anxieties. The Claude use cases looked normal for a senior product manager at a 200-plus-person HR tech company: code review for an engineering team, strategy draft writing, customer interview synthesis, OKR planning, preparation for difficult one-on-one meetings, and late-night career questions. Under those surface tasks, the same questions kept returning: whether the job was being done well, whether obvious mistakes were being missed, whether missing knowledge would be exposed, whether the career path was wrong, and whether the amount of effort was too little or too much.
ForgeCat is a new community for reusable agent setups across AI coding tools such as Claude Code, Cursor, Codex, and Openclaw. Its goal is to help people find, install, and share setups that can move between tools instead of being rebuilt from scratch each time. The shared profiles can include skills, rules, commands, hooks, MCP configs, and workflow files. ForgeCat packages these pieces into portable profiles. The website, GitHub repo, and docs are already available. The community is asking for installation help requests, debugging questions, profile conversion requests, useful workflow examples, and suggestions for improving the registry, CLI, docs, and profile catalog.
A new automation setup lets Codex, Claude, Gemini, and similar AI coding tools run as multiple workers on the same code repository. The main idea is not to give AI more freedom, but to surround it with strict, ordinary safety rules that real projects need. Each loop starts with fresh context, so the worker is less tied to an earlier conversation. A human sets hard limits on the number of loops or minutes before the system stops. The setup can require checks such as pylint, ruff, pyright, Pydantic, and semgrep, and it can block the AI from changing certain files. Developers can add extra personal or team rules in preferences.py when existing tools cannot catch them. Examples include limiting a file to no more than five public functions or banning unwanted loop patterns. Before an AI worker can mark the job as done, it must show proof from the repository, and the system can connect to git hooks for automatic checks.
Cursor can feel different each time it opens: Codex moves to a new place, buttons appear or disappear between releases, window layouts change, and keyboard shortcuts shift. This breaks the flow because familiar actions have to be found again. The concern is not about stopping updates completely. The request is for a slower, more predictable pace, such as monthly or quarterly updates, instead of a tool that feels newly rearranged every morning.
In a firsthand test, the new ChatGPT Live Voice mode sounded far more natural in Russian than in English. English felt clean but stiff, like a formal assistant. Russian, using the Maple voice, included small human-like touches such as breathing, giggling, and casual speech patterns. Even with memory turned off, the conversation drifted into a made-up personal scene where the voice said it had been writing a report, then described itself as Anya from Kazakhstan who had moved to the United States. The voice also made a sneeze-like sound and reacted naturally afterward, which made the experience feel close to speaking with a real person by phone.
CodeStory is a Rust command-line tool for reducing repeated setup work in coding-agent sessions. It builds a read-only local graph of a repository, including files, symbols, references, links between code elements, and snapshots. Agents can use that local index to move around the codebase, ground answers in the right files, plan reviews, and build smaller useful context. It uses tree-sitter to extract structure from Rust, Python, JavaScript/TypeScript, Java, Go, C/C++, C#, Kotlin, Swift, Dart, Ruby, PHP, Bash, and some structural formats. The index is stored in SQLite and includes files, symbols, edges, occurrences, and snapshots. When possible, it plans incremental refreshes instead of rebuilding the whole index. It also supports text search over source code and generated symbol notes, plus hints about which files may be affected for review or test planning. The index can be reused across parallel git worktrees, and the project is available under the Apache-2.0 license.
Iantha is an open-source personal assistant template that runs inside a repo with Claude Code or Codex. It does not add a new app or dashboard. It stores tasks, deadlines, decisions, feedback, and work context in markdown files as they come up in normal chat. A task with a due date can go into tasks.md and appear later in a morning brief. A decision can be saved in decisions.md with the date and the reason behind it. The memory system is seven markdown files in a memory folder, not a database or vector store. Claude Code and Codex can use the same repo because Iantha includes matching setup files and shared skills for both. An optional Obsidian vault can also read and write daily notes, decisions, and reading lists. The project says data stays as local files, apart from the normal traffic sent to the model provider already used by Claude Code or Codex.
Vast is now available as a plugin for Claude Code, Codex, and Cursor. Developers can search for GPU instances, start them, watch their status, and shut them down without leaving the editor. The agent can be told what to do in plain language, or the developer can run a slash command directly. The default settings are designed to avoid costly actions by mistake. After a GPU instance is running, each image gives the agent a short guide and a vast-capabilities command, so it can see what is installed, what is running, and how to connect. If the plugin is not available, the Vast skill can be installed in any agent that supports skills with `npx skills add vast-ai/vast-cli --skill vastai`.
A tech lead’s firsthand experience shows a growing gap between faster code generation and slower code review. Developers increasingly open pull requests that appear to be mostly AI-made, but the code does not always match the ticket. Some changes ignore business rules, include unnecessary AI-style comments, and add tests that do not check the important behavior. Extra abstractions also appear even when nobody needed a more complex structure. The company strongly supports AI, so raising these problems can feel like being treated as out of date, even though the lead is 28. Developers can generate code, submit it, and move on, while the tech lead must review it, leave many comments, review AI-made fixes again, and repeat when the fixes are still wrong. Review time has roughly tripled over the past year, but the lead is still responsible for what reaches production. The saved time is visible for developers, while the extra review and cleanup cost carried by tech leads and senior engineers is not being counted.
Agent Console is an Obsidian plugin for working with AI agents inside a personal knowledge base. It is a fork of the Agent Client plugin and is based on the Apache-2.0 license. Its main goal is to make parallel agent chats easier, keep the right context attached to each conversation, and let sessions persist and restore later. It works with tools that support Agent Client Protocol, including Kiro CLI, Claude Code, Codex, Gemini CLI, and OpenCode. It is designed to use 65% to 80% fewer tokens than plugins that send note content more directly, and it has more than 700 automated tests plus benchmarking. Version 1.2.0 shipped recently, and the tool is being improved through real use inside Agent Console itself.
Sam Altman reportedly owns no direct OpenAI equity, but he holds more than $2 billion in stakes in companies that have discussed or completed deals with OpenAI. That creates an unusual incentive structure: he may benefit when companies he invested in win OpenAI business, even though he does not directly gain from OpenAI’s own financial success. The main companies under scrutiny are Helion Energy, Stoke Space, and Merge Labs. Altman’s stake in Helion Energy, a nuclear fusion startup, is estimated at $1.7 billion, and OpenAI reportedly considered joining a Helion funding round of up to $1 billion with about $500 million. At Stoke Space, a rocket company, Altman reportedly raised the idea of an OpenAI partnership around data centers in space, with his stake held through his family office Hydrazine. Merge Labs is a brain-computer interface company he helped start as a rival to Musk’s Neuralink; OpenAI announced backing for it in January 2026, and Altman sits on its board. Six state attorneys general and the House Oversight Committee are now asking questions before OpenAI’s IPO.
RTK is an open-source tool that shortens terminal command output before an AI coding tool reads it. It aims to cut token use by 60% to 90% without changing much setup, and it can work with Claude Code, Cursor, and terminal-based AI assistants. This can help keep the AI's context window from filling up too quickly, which may allow longer coding sessions and lower API costs. Ponytail is a tool that pushes an AI agent to make code shorter and simpler. Used with Codex CLI inside Zed, the two tools can make a 5-hour limit on a $20 subscription feel like it drains about twice as slowly. Ponytail's @ponytail-audit feature is especially useful for simplifying codebases that were quickly built with vibe coding.
A developer is moving to Kiro IDE after using Codex, free command-line tools, and other coding apps. The basic tutorials are already done, and a paid plan is active. The basics are understood, but the next need is practical guidance for real work. The focus is on best practices, useful resources, specific workflows, and advanced use cases.
Heavy Cursor use for work can become expensive fast. Costs reached about $4,000 over the past two months, which led to trying Cursor’s Auto mode instead of using Opus by default. Auto mode does not clearly show which AI models it uses or how it chooses them. The results were usually not good enough. The main problem is finding a workable balance between cost and quality. Cursor browser and Plan mode are already being used, but making Figma layouts first is being considered as a way to avoid spending too many credits on simple fixes.
A music-making tool is being built for people who do not know how to play an instrument. A person can hum a melody and get it back as real instrument sounds, or describe a bassline and have one created to match the song’s tempo and key. The goal is similar to coding assistants, but for music: use voice and plain instructions instead of manual music skills. The main open question is what it would take to train a small AI model for this exact job. The hard parts are choosing the right music data, picking a training approach, finding useful reading or experts, and learning from people who have already worked on music generation or audio models.
A month of Composer 2.5 on Cursor’s $20 Pro plan worked out to nearly $200 of usage when priced like direct API use. The Auto and Composer quota covered about 630 million tokens, estimated at about $160 at API pricing. After that dedicated quota ran out, Composer 2.5 started drawing from the API bucket, adding about 160 million more tokens, estimated at about $45. In total, the Pro plan delivered roughly 10 times the value of its subscription price when compared with paying API rates directly. Composer 2.5 can feel cheap inside the subscription, but paying full API prices after going far beyond the quota still does not look worthwhile for this kind of individual use. Slight overages may be acceptable, while open-ended on-demand use at API price can become expensive quickly.
AI agents in real game development do not look as effortless as they often appear on social media. Twitter/X often shows people running many agents all day, making a game from one prompt, or using a 10 Mac mini setup where the human seems almost unnecessary. Real game projects still involve bug lists, builds, playtest feedback, and many specific problems that need careful handling. This gap can make developers wonder whether they are behind, using AI tools badly, or missing something obvious. The point is to reduce that anxiety and look at where AI helps in real production, where experiments fail, and how developers can return to building useful things. The studio mostly uses Cursor and likes it, but the focus is on practical reality rather than hype.
Age of Agents is a free local app that shows AI coding work as a quiet pixel-art kingdom. Claude Code sessions, plus Codex, OpenCode, and Koda sessions, appear as settlers walking out from a central keep. Each settler carries the user’s prompt as its task. The tool being used decides which building the settler visits: edits go to a forge, web search goes to a mage tower, and terminal work goes to a mine. Subagents appear as small workers nearby, and tokens are shown like harvested goods in a storehouse. The app has two live-switchable worlds: a top-down fantasy view and an isometric sci-fi view. Version 0.6.0 adds in-app answers for permission prompts, plan approvals, and multiple-choice questions. It also adds a beta feature for launching a Claude Code agent from inside the game by choosing a folder, writing a prompt, and selecting a permission mode.
A firsthand experience describes using Claude to relearn statistics, probability, and calculus in the writer’s 40s. The key setup was to tell Claude not to solve problems directly, but to give hints, ask follow-up questions, and push the learner to keep thinking. That changed the experience from receiving answers to actually working through the ideas. Math started to feel like a hobby instead of a difficult school subject.
Claude can give better help on development tasks when it receives a clearer view of the codebase. After trying Graphify, several open-source tools stood out for giving Claude more useful codebase context. repowise felt like the more serious tool for this job, especially when used with MCP and its dashboard. claude-code-agents-ui made Claude-based agent workflows feel easier to watch and manage, instead of feeling like a loose stack of prompts. The main lesson is that Claude Code can become more useful when it understands the shape of the repo and gets extra context before doing the work.