Talk to the agent without typing "agent" first — just say what you want to build. The agent now has an animated face that shows what it's doing, and you can give it a name in Settings. Every provider's API key field links straight to where you get the key. Added AirLLM so you can use a large model running on your own GPU over Wi-Fi, and fixed local models being cut off mid-reply. Markdown formatting in notebooks now applies properly. What iOS users will notice 1. Just say what you want. No more typing agent first — a sentence goes to the agent, ls and git status still run as commands. agent … still works. 2. The agent has a face and a name. An animated avatar shows whether it's thinking, working or waiting on you. Name it in Settings and it answers to that name. 3. Get API key links. Every provider's key field now links straight to that provider's console, and Settings says plainly that you pay the provider directly. 4. AirLLM support. Point the app at a big model running on your own GPU machine over the LAN, alongside the existing Ollama and MLX-LM options. 5. Markdown toolbar works in notebooks. Bold, italic and headings applied and then vanished; they now stick. 6. Local models stop timing out. A model running on your own hardware is given time to answer instead of being cut off at two minutes mid-reply.

Choose How You Want to Build When creating a new project, iCoderAgent asks how you want to work: • Vibe Coding — Describe what you want and the agent builds it. This is the existing experience and remains unchanged. • Spec-Driven — The agent interviews you first, creates an agreed specification, and only then starts coding. This choice is made once per project, remembered, and cannot be changed later. Spec-Driven Interview The interview is a structured form, not a chat stream. Each specification document appears as a card. Questions are presented with selectable options wherever possible, with free-text input when needed. The interview captures the project's features, workflows, UI, data, integrations, security, constraints, project structure, and acceptance criteria. The resulting specification becomes the source of truth. Specification Is the Gate In Spec-Driven mode, the agent cannot write code simply because it thinks code is needed. Every change must be linked to an approved requirement through an open task. Before creating or modifying a file, the agent must confirm: 1. The file is on the agreed file list. 2. It is in the agreed folder. 3. The change belongs to an open task. 4. The task cites the relevant requirement. 5. The change does not introduce unapproved functionality. If a new file is required, it must go through the approval process: New file → Specification update → Agreement → Task → Implementation The same applies to new features: New feature → Specification first → Agreement → Task → Code Acceptance Criteria Must Be Checkable Acceptance criteria must be objective and verifiable. "The app is fast" is rejected. "Dashboard loads within 200 ms after the required API response is available" is accepted. Criteria should be specific, observable, measurable where appropriate, reproducible, and unambiguous. If a criterion cannot be objectively checked, the agent asks for clarification rather than making assumptions. Done Means Verified In Spec-Driven mode, Done means verified, not merely implemented. For every acceptance criterion, the agent provides a command, test, or executable check that proves it. Example: Requirement: Dashboard response time must be below 200 ms. Verification: npm run test:performance -- dashboard The project runs the verification. Pass → Verified Fail → Not verified A failing command can never mark an item complete. Some requirements may require visual or manual inspection. These are recorded as Inspected, not Verified. Each criterion therefore has one of four states: • Verified — proven by a successful executable check. • Inspected — reviewed manually but not machine-proven. • Failed — verification was attempted and failed. • Pending — verification has not yet been performed. Gate Audit Trail Every decision made by the Spec-Driven gate is recorded in: .icoder/gate.log The log is accessible from the Files tab and records project mode selection, specification decisions, approved requirements, file/folder decisions, tasks and requirement references, specification changes, verification commands and results, and decisions that allowed or blocked implementation changes.

Plan before you build Ask for an app and the agent now agrees what it's for first — who it's for, what it should do, and what it won't do — then writes that to goal-spec.md in your project. It designs the look with you too: three real mockups to choose from, not a surprise at the end. The specs stay in step as it builds, so what you agreed and what you got don't drift apart. You can check the result against them at any point. Security warnings as it writes Every file the agent writes is checked as it goes — an API key left in a config file, a query built by string formatting, a password sent to a log. It's told immediately and fixes it. No setup, no extra cost. Files tab Save, Auto Save and Save As in the editor menu, so you can keep a copy under a new name or let it save as you type. Add Files… brings a file you already have into the project. Settings The model list and key status now collapse, so the things below them are reachable without scrolling.

Smarter answers, fewer tokens Ask a question and get the answer in the format you need. Enable TypeSafe in Settings to let iCoderAgent use a specialised decision model for choices, scores, and true/false questions—returning a clear result with a probability instead of a long explanation. This can also reduce usage of your main AI model. Example: Is Python dynamically typed? → True (99%) Pick one: Swift | Kotlin | Dart → Swift (98%) Open Jupyter notebooks anywhere Open .ipynb files directly from Files, Mail, or iCloud Drive and iCoderAgent will launch them ready for you to work with. Install specialist agents Add specialised AI agents from a catalogue. Simply point the plugin installer to a folder containing agent roles and choose the ones you want to use. Better links Notebook links now open where you expect them to. Opening a notebook link starts a new notebook instead of unexpectedly taking you back to an existing notebook. Latest OpenAI models iCoderAgent now supports the latest OpenAI models, including the newest reasoning models, giving you more options for advanced coding and problem-solving. Bug fixes & improvements Fixed an issue where opening a notebook from iCloud Drive could cause a cell's output to disappear when saving. Fixed the Usage Report showing an incorrect “Unknown” model. Additional stability and usability improvements.

Notebooks • Open, edit and run Jupyter/Colab .ipynb files right on your phone — or on a connected laptop or cloud host for the heavy work. • New "Machine Learning notebook" project type. • Ask AI on any cell: describe what you want and the code is written for you. • Shell magics work: !pip install, !command, %cd, %%bash. • Add a cell where you are, not just at the end. Line numbers, full-screen editing, rich markdown formatting (font, size, colour, alignment). Python • Per-project virtual environments that behave like the desktop ones — create, activate, deactivate, and pip install packages on the device. • Starlette and FastAPI apps now run in Preview, alongside Flask. Node.js • Node templates run on-device: npm start/test, servers in Preview, and a React template that bundles on the phone. Themes • Per-account themes, terminal fonts and text sizes, switching live across every screen. Light themes are properly readable everywhere. • macOS-style window controls throughout. Projects and files • Projects live in a folder you choose, and the app shows you where it is. • Two-way sync with a connected host now pulls changes back to the phone. Fixes • Crash reports can be diagnosed by the AI and turned into a reviewed bug report before anything is sent. • Many light-theme contrast fixes, help tables that line up, and Preview chrome that follows your theme.

• New Support section in Settings: send feedback, report a bug with optional diagnostic logs, and open the Help Center. • Refer a friend: share your code or QR link from Settings → Support. When your friend joins and completes their first conversation, you both get 10 bonus conversations. Earn rewards for up to 50 friends. • Have a referral code? Redeem it on a new account under Settings → Support → Share iCoderAgent. • Fixed conversation counting: only successful replies count toward your quota, and each one is counted once.

LONGER SESSIONS • Fixed: a long conversation could stop working entirely. Once a chat grew past a certain size, every following message failed with an error and the only way out was starting over. • Fixed: the agent no longer stops early on tasks with repeated steps — like taking a screenshot of each screen — which it previously mistook for getting stuck. • Fixed: long runs that produce a lot of screenshots no longer run out of memory, and chat cards no longer go blank. • Jump to the start or end of a long conversation with the new buttons in the corner of the terminal. NEXT STEPS • Tapping a suggested next step now shows the agent's full working detail as it goes, instead of just a progress bar. • Its result is properly formatted, and offers its own follow-up steps — so you can keep tapping through a piece of work. TASK LISTS • The task list now finishes properly: it ticks everything off and says so, or names what was left undone. • A completed list stays on screen so you can see what was done, and clears when you start something new. PROJECTS • Switching projects now resets the chat to that project's own conversation, instead of leaving the previous project's messages and task list on screen. • New and opened projects get their standard files set up consistently. USAGE AND COST • Token and cost totals now clearly show which project they belong to — each project tracks its own. • New "By Project" breakdown in the usage report, plus /project_usage_report in the terminal for a quick summary. REMOTE AND SSH • Fixed: an SSH command that failed to start would hang with no error. It now reports the failure. • Fixed: usage syncing with the relay server.

Project wizard — start a new project the way Xcode does, with bundled templates for 14 stacks across Apple, Android, web, Python, Rust, Go and .NET. Files, rebuilt — a real tree view, create files and folders anywhere, share any row, show hidden files, and switch or close projects from the project bar. Opening a large project no longer freezes the screen. Chat per project, per day — conversations are saved with the project and picked up where you left off. Next steps you can tap — the agent's follow-ups are now buttons that run the action, not text to retype. Task lists that finish — the checklist now closes out when a run ends, so you can tell done from abandoned. Fixes: long runs no longer blank the chat or run out of memory, .icoder stays out of your git remote, longer answers are no longer cut short, and a disconnected laptop no longer burns a whole turn.

In this update: • Sign in with Apple • Preview tab is now a full in-app web browser • Run long builds and tasks in the background, with a notification when they finish • See a “running” indicator from any tab while the agent works • Improved in-app help and documentation • Bug fixes and stability improvements Shorter alternative if you prefer minimal: • Sign in with Apple • Preview tab is now a real web browser • Background tasks with finish notifications • More accurate audio/video transcription • Bug fixes and improvements