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.
Most existing Telegram MCP tools rely on the Bot API, so they cannot read personal chats. mcp-telegram-local is a local Telegram MCP server that runs on the user’s own computer. It uses MTProto instead of the Bot API and connects to a personal Telegram account through QR login. The goal is to let AI tools such as Claude access personal Telegram conversations, not just bot messages. The code is available on GitHub.
Argus Red argues that public AI models from Anthropic and OpenAI are built to refuse offensive security tasks, while specialized cyber models are usually limited to enterprise customers. That can leave small and mid-sized companies without strong AI help for finding serious weaknesses. Many current AI security tools still sit on top of general models, so they inherit the same refusals. This tool uses a model post-trained on ten years of capture-the-flag contests. It will not be open to everyone, but the goal is to make this kind of tool available to responsible smaller companies, not only large enterprises. It runs through a CLI with two modes. Security scan is a read-only check of a local codebase and only reports issues it can connect to a specific file and line. Pen test is an active mode that tests systems from an attacker-like angle.
A solo developer can build high-quality apps but only has a basic grasp of marketing. Building an interesting app does not automatically mean people will discover it. The practical question is whether Claude can help with marketing through an agent, a pipeline, or specific working methods. The goal is to use Claude to fill the marketing gap around a product made by one person.
Major AI observability companies sell products that let agents review their own past work logs, spot errors and inefficiencies, and improve automatically — a technique called a self-improvement loop. These products are expensive, hard to adopt, and keep your data on their servers unless you pay enterprise prices. The underlying mechanism is an LLM acting as a judge that evaluates another agent's behavior. A developer reasoned that anyone with a Claude Code subscription already has what it takes to run this locally, and built Kyoko, releasing it as free open-source software. Kyoko has four parts: local trace collection and storage using the Open Telemetry standard, an analysis step that spins up a local Claude Code instance to read through those traces, an evaluation harness that runs code and LLM checks to track performance over time, and a human oversight dashboard where you approve any fixes before they are applied. Everything runs on your own machine; you control your data. It can be driven by a human through the dashboard or operated automatically via a command-line agent. In a test using Sierra's Tau2 benchmark, the same underlying model achieved a 24% performance gain after going through Kyoko's fix cycle.
A small private sauna rental business in Poland moved from an outdated Wix site to a custom system built with Claude. Before this, bookings came through WhatsApp, pricing was not clearly managed in one place, and someone had to turn on the heater by hand about an hour before every session. With limited technical background and no history of shipping a real web app, the maker spent several months building a customer site, an admin dashboard, and heater automation. The customer site was rebuilt in Next.js with booking, pricing, FAQ, Google reviews, and Polish and English pages. It was deployed on Vercel, with automatic updates when new code is pushed. The private admin dashboard lets the parents manage bookings, clients, live sensor readings, and a calendar. A Raspberry Pi in the sauna building reads a temperature sensor, controls the heater relay, checks the booking schedule, and starts preheating when a booking is coming up.
In late June 2026, Gemini users reported that usage limits were being consumed much faster than before. Simple prompts, normal chat, and basic transcript cleanup appeared to use up large portions of the quota, sometimes making the service feel limited every half hour. One EU user said a single simple request, with no document or coding task, used 7% of a 5-hour allowance; the same person estimated that the limit fell from about 50 prompts per 5 hours to roughly 14. Another user said the same prompt used to take 5–7% but now could take 12% or more, even in a new chat. Gemini Plus with Flash Extended also drew complaints because regular conversation, not coding, could hit the limit in about two hours. Some users also reported official Google policy warning emails, but those warnings appear more related to abuse-protection rules than to the quota change itself. At the same time, Google Antigravity seemed to use separate limits from web Gemini in at least one coding workflow, with almost no visible quota drop after use, so coding-tool limits may feel very different from chat limits.
Claude Code 2.1.191 is a practical cleanup release for people who use the tool inside a terminal. It adds `/rewind`, so a conversation can be resumed from the point before `/clear` was run. It fixes a scroll bug that pulled the screen back to the bottom while earlier output was being read during a streaming response. It also fixes stopped background agents coming back after they had been stopped, and it now gives a clearer message when `/voice` is disabled by an organization policy. Several terminal-specific issues were corrected, including a truncated `/login` link in Windows Terminal and Cmd+click links in Ghostty when used through ssh or tmux. Agent views were cleaned up so built-in slash commands such as `/usage` are not sent as prompt text to background sessions, and pasted images show as `[Image #N]` instead of long filesystem paths. Nearby 2.1.x releases show the same direction: stronger sandbox controls for credentials and secrets, clearer organization model restrictions, command-line MCP login, better auto mode denial messages, hook matching fixes, and an option to disable mouse clicks in fullscreen mode.
Flama 2.0 adds tools for downloading a large language model, packaging it into a .flm file, running it on a local computer, and serving it over HTTP. A model from Hugging Face can be fetched with `flama get`, which saves the model data and settings together. `flama model run` sends one prompt and returns the full answer, while `flama model stream` shows the answer as it is being produced. Settings such as temperature and maximum token count can be changed, and a JSON list of messages can be passed in for multi-turn chat testing. `flama serve` starts a local server, with the example running on `127.0.0.1:8000`, and provides both an API and a built-in chat screen. It supports OpenAI, Anthropic, and Ollama-compatible connection styles, so existing tools can call the Flama server like a normal AI API. Claude CLI can also be pointed at the local Flama Anthropic endpoint by changing `ANTHROPIC_BASE_URL` and setting the default model names to a local model such as `gemma`.
A solo maker built a food journaling app with help from Claude and GitHub Copilot. Claude was used for market research, and feedback also came from people nearby and a few dietitians. The app focuses on food journaling and does not include a calorie counter. Before launch, Claude and Gemini keep giving very positive reactions when asked to validate ideas. That creates a problem because weak spots, market risks, and reasons people may not use the app are harder to see. Finding the right people to ask is also difficult without spending much time on social media.
The goal is to run a local AI on personal devices and feed it chosen books and data. It should have several roles. As a teacher, it would use textbook PDFs to help with learning. As a house manager, it would track food, bills, where things are, and notes to remember later. As a personal chef, it would use recipes and diet needs to create detailed meal ideas. Because this setup would contain sensitive personal information, the key requirement is that it stays fully disconnected from the internet.
A solo indie developer named Clem released SubjectiveZero, an open-source agentic node editor for creative coding (visual and interactive work). The tool tries to solve a common design tradeoff in creation tools: low-level tools (shader composition, raw code) are powerful but hard to learn, while high-level tools are easy to start with but users eventually hit walls on edge cases. SubjectiveZero lets users start with abstract prompts describing their goal, then move toward more specific prompts or directly edit the generated code as needed. The idea is to treat the abstraction level as a slider the user can move in either direction. The underlying agent orchestration is built to understand the user's context so it can translate intent into working output.
OpenAI introduced Deployment Simulation, a way to check how a new model may behave before users get it. The method takes earlier conversations, removes the old model’s answer, and has the release candidate answer in the same place. OpenAI says it removed account-linked and personally identifying details, and only used ChatGPT conversations from people who allow their data to be used for model improvement. The study looked at about 1.3 million de-identified conversations from GPT-5 series Thinking models between August 2025 and March 2026. It predicted the real-world rate of 20 unwanted behaviors, such as tool-related deception and disallowed content. Compared with hard test prompts, the method better predicted whether these problems would rise or fall after release, with a median error of about 1.5 times. It also reduced the chance that models could tell they were being tested. GPT-5.2 marked real user traffic as test-like 5.4% of the time and simulated traffic 5.1% of the time, while synthetic test sets were flagged as tests almost 100% of the time. OpenAI also tried the method in Codex-like coding agent settings, where tool calls and repository state must be simulated to make the test feel close to real use.
Cascade is a neural image codec that stores a very small version of an image and uses a generative image model such as Gemini to rebuild it. It was trained on 1 million images and focuses on familiar objects and shapes. Its goal is to keep most of the original style, color, and structure while using far less data than JPEG. On test images, its average bitrate is 0.03, with a range from 0.005 to 0.16. The system uses three parts: a VQ-VAE, a cascading conditional convolutional network, and a generative image model with a LoRA. Gemini is the default decoder, while SDXL is available as an offline option. Setup means cloning the GitHub code, downloading models from Hugging Face into a models folder, setting GEMINI_API_KEY, then running encode.py and decode.py to create and restore .nit files. Options include qp for the quality-size tradeoff, tokens-only to skip model or API use, and n-decode to generate several outputs and keep the one closest to the guide.
An ebook-to-audiobook service was built to turn ebooks into long-form narration with realistic AI voices. Subscription-based voice tools felt too expensive for limited use, while the open Kokoro voice model sounded natural enough for hours of listening without the tired feeling caused by older robotic text-to-speech voices. Kokoro is designed to run fast, but generating long audio on a 12-core laptop CPU was still too slow. A cloud GPU setup was chosen so audiobooks could be generated quickly for a self-hosted library and possibly offered as a product for others. The project had two goals: build a text-to-speech product focused on ebook conversion, and get real experience with multi-agent AI coding workflows. Almost all of ebookaloud was written by DeepSeek v4 inside OpenCode. Over one month, the build used about 750 million tokens and cost $12 in credits. Each feature or change followed a plan, implement, test, review, correct, and commit cycle, using a mix of Pro and Flash agents.
A Claude Code orchestrator that manages other agents can become messy if it only relies on the model’s default judgment. Each orchestrator needs a mission document that says what it should do, how it should do it, and when it should ask a person for help. For a bug-report workflow, the orchestrator can check email for customer feedback, create a local ticket, send a child agent to reproduce and fix the bug, and alert the maker when the work is ready for QA. This lets the person stay out of the process until the review step. The orchestrator also needs a loop, because a one-time run is closer to a script than a continuing helper. Re-reading the mission document every 10 minutes or every hour gives it a steady rhythm for checking what needs to happen next.
A mortgage application required 17 separate PDF files, which made checking and tracking documents tedious. Merging everything into one long PDF reduced file switching, but it also made it harder to see where one document ended and another began. The solution was a two-dimensional viewer: scroll sideways through pages inside one file, and scroll vertically to move between different files. The format adds metadata to a PDF so document boundaries are stored inside one combined file while still staying compatible with normal PDF handling. The new format is called .pdfx, but it can also live inside a regular PDF with metadata. Claude helped build about 80% of the project in roughly 2 hours, shortly before access to the most advanced model ended. The app is open-source, built with Electron, includes some Liquid Glass-style interface work, and still has rough edges.
echo•mux is a personal tool for playing music on several Bluetooth speakers from different brands at the same time. It avoids buying a new Wi-Fi speaker system such as Sonos and instead reuses existing Edifier, Sony, and Sackit speakers. The main challenge is that each speaker can play sound with a slightly different delay, so the system needs per-speaker latency alignment. Raspberry Pis act as multi-room Bluetooth hubs, controlled through a mobile-first web UI. Spotify Connect support is part of the setup, so music can be started from Spotify. Extra Bluetooth antennas costing about $10 each were added to improve coverage, including outside in the backyard. The project was built over two weeks in spare time with Claude, which also made test-driven development easier to follow.
The workflow would let Hermes react to bug emails, check Sentry, try to fix the issue, and open a pull request. The coding work should run through Codex CLI, Codex 5.4, and an existing set of personal skills and tools, because that setup is already trusted for code quality. The main problem is how Hermes can hand work to another tool or model in a reliable way. acpx was tried one or two months earlier, but it was unstable and eventually stopped working. The current experiment is a Hermes skill that calls Codex CLI directly with a command like `codex exec ...`. The practical need is advice from people running real products with external users, where this kind of handoff has to be dependable.
Cursor’s new features can be hard to understand from the official documentation alone. The example raised is “MultiTask” mode in the IDE agent window. The feature name is visible, but the available docs do not clearly explain what it does, when to use it, or what result to expect. This leaves users learning by trying the feature and seeing what happens. There is also some optimism because Cursor has reportedly hired people who could help improve this area.
A Cursor Enterprise admin built a system that watches Cursor’s official docs and updates internal Cursor guides when something meaningful changes. A Python script runs on a schedule in GitLab, pulls the raw markdown files from cursor.com/docs, and compares them with the last saved version. When it finds a real factual or workflow change, it sends a webhook; marketing copy changes are ignored. The webhook starts a team-owned Cloud Agent through Cursor Automations. A strict prompt maps specific Cursor documentation pages to the right internal admin and user guide files. The agent reviews the change and prepares updates to the internal docs. The updates are opened as merge requests in GitLab, then mirrored as read-only copies into Azure DevOps repos and project wikis. The internal docs cover governance, admin guidance, and standard user guidance.
In firsthand use, Codex Plus and Claude Pro cost the same, but Codex Plus appears to last much longer before hitting usage limits. Claude Pro feels much more restrictive for sustained work. Claude is still hard to drop because it performs well for web design work and Claude Design tasks. The practical question is whether the $100 Claude Max plan gives enough usage to replace both Codex Plus and Claude Pro. The workflow is mainly development through Claude Code CLI and Codex, not the regular chatbots.
TBD is an open-source Mac tool for managing multiple coding agents. It has a visual app, but its main rule is that every action available in the app must also be available through the command line. That makes it easier to automate work or let agents control the tool themselves. It is meant to work well with tools like agent-channels when agents in different worktrees need to talk to each other. The tool grew out of long daily use of Conductor, where small unresolved problems became hard to tolerate in such a central workflow tool. dmux, claude-squad, and agent-deck were also tested, but heavy keyboard shortcut use and tmux-style workflows did not fit the need. TBD started in March 2026 and has become stable enough that several current and former colleagues now use it as their daily tool.
A 17-year-old developer is trying to choose a long-term path in software. They have programmed for about 5 to 6 years, starting with Discord bots and later trying web development, IoT, networking, and automation. From the beginning, they rarely wrote everything from scratch; they mostly edited existing code, copied and adapted examples, built on starter templates, and later leaned heavily on AI assistants. They are comfortable with Next.js, Three.js, Python, C++, and learning new tools, but they do not memorize syntax. Their strength is reading code and understanding what it does, even across different languages. They tried Framer and frontend work as a way to make money, but found that pixel-level visual detail and UI/UX work do not interest them much. They enjoy system design, connecting parts together, backend architecture, and IoT more than frontend styling. Their long-term goal is to become a strong software engineer and eventually build companies, not just freelance websites, but they are unsure whether to go deep into backend and distributed systems, focus on IoT platforms, or keep exploring.
A neuroanesthesiologist rebuilt a hospital fellowship website with Claude and Claude Design over one weekend. The site had been neglected for about two years, had outdated information, and received roughly 8 visitors a month. Hiring a web developer would have meant writing a brief, waiting weeks, and likely getting a generic template, so the work was tried directly with AI instead. The doctor supplied the real content, including the curriculum structure, medical context, program details, and a dark, clean visual direction. Claude handled the HTML/CSS, responsive layout, and repeated changes. This was a landing page, not a complex app, but the important shift was turning a clear idea into a live, decent-looking site without a middleman. Asking for the desired result worked better than describing exact layout details. A request like making a section read like a prospectus rather than a brochure was more useful than trying to control the design piece by piece. After three months, traffic was about 14 times higher.
Claude is being used to move years of university and law school notes into an Obsidian personal knowledge database. The goal is to review every lecture, assignment, and writing sample, then create detailed notes for each concept. The instructions are long and specific, and they also say the work can take multiple sessions and does not need to be rushed. Claude performs well for about half a class, then slips back into making one broad summary page instead of breaking the material into separate concept notes. The work has to be stopped and corrected with firm reminders before Claude returns to the requested format. Claude then acknowledges that it took the easier path by leaning on outlines instead of turning the notes into smaller concept pieces. The core problem is that a long AI task can start correctly but lose discipline over time and drift toward shorter, less careful output.
Cursor’s recent changes point toward keeping more of the software-building process inside Cursor. Its autonomous model, Origin, and mobile app bring editing, running code, version control, and the model used for coding help closer under one brand. The upside is clear: fewer app switches, one bill, and a smoother daily workflow. The risk is that each added piece can make it harder to leave later, even if no single change feels restrictive at first. The main concern is not editor features or Git integration by themselves. The bigger concern is a future where the model used to reason about code becomes hard to swap out. Keeping a second workflow in Verdent, CLI, and Git helps preserve the ability to work outside Cursor if prices, policies, or product direction change.
In Cursor 3.9.8 on macOS, custom MCP servers appear connected in settings but are not available to the chat agent. Servers listed in `~/.cursor/mcp.json` show as enabled with a green status in Tools & Integrations, and their tools are visible there. When the chat agent is asked to use one of those tools, it says the server does not exist. The only servers the agent lists as available are plugin-based MCP servers such as `context7`. Restarting Cursor, reloading the window, starting new chats, toggling the server, and removing and adding it again did not change anything. The servers run correctly on their own and expose their tools, so the likely issue is that Cursor launches them for settings but does not pass them into the actual chat agent session. The project MCP descriptor folder also only shows plugin-based servers, while the `mcp.json` stdio servers are missing.
Claude may count a retry after a server error as if it has to read the whole context window again. With a 200k context, even a simple retry can use a large part of a 5-hour usage limit. Small prompts can end up consuming around 40% of the allowance when repeated errors force several retries. The lost usage comes from a platform problem, not from a mistake by the person using the tool. Support can feel cold when the answer is that nothing can be done unless the customer is on the highest-tier plan. Possible fixes include usage credits, refunds, or another way to restore usage lost because of service errors.
Claude Code runs the main development workflow, while Codex handles selected review steps for risky changes. Running several Claude reviewers can look like strong agreement, but models from the same family can share the same weak spots and miss the same bug. This workflow keeps Claude as the coordinator and sends narrow, adversarial checks to Codex. Codex is called through codex exec and returns results in a strict JSON format. The goal is not to guarantee correctness, but to reduce repeated mistakes and shared blind spots that can happen when one model family reviews its own work. There are no benchmark numbers yet, so this is best treated as an advanced quality-control pattern for debugging, context management, subagents, and multi-agent workflows.
SenseNova Office Skills uses a separate SKILL.md file inside each skill folder, so an agent can read the instructions while it is working and understand what each skill does. This replaces a setup where tools are mainly defined as function calls or fixed API shapes. The claimed benefit is that the agent can discover skills, understand them, and connect them together without hardcoded routing. The skills are arranged in layers. The lower layer, sn-image-base, exposes basic model abilities such as image generation, recognition, and text improvement. Higher-level skills such as sn-infographic combine those basic abilities into full workflows, including automatic prompt scoring, multiple generation rounds with quality ranking, and choosing from 87 layouts and 66 styles. Other listed skills cover image imitation from reference images, resume image generation, PowerPoint generation, Excel data analysis, and a deep research agent.