What's New

Smoother to read, easier to follow.

Enhanced in 1.21.2

• Long prompts fold away. Paste a long brief and Naseem shows its first lines with Show more, so the reply stays right below it.
• The Files pane stays current. Files added, renamed or deleted — by Naseem, Finder, git or a build — show up on their own, no Refresh needed.
• Choose how hard Claude Haiku 5.5 thinks. Haiku 5.5 now has the Effort menu, from Low to Max.
• Easier to see that Naseem is working. The “Working…” line under a reply is larger and clearer.
• Links feel like links. Show more, Open… and other links show the pointing hand when you hover them.

Quick answers when you want them.

Enhanced in 1.21.1

• Quick answers when you want them. In the Effort menu, GLM, DeepSeek, Kimi and MiniMax models on Ollama and Ollama Cloud now have an Off option. Switch it off for fast, simple replies, and back on when a task needs real thinking.

Edit your files without leaving Naseem, and smoother testing in the browser.

New in 1.21.0

• Edit files right in Naseem. Spotted something to change in the Files pane? Click the pencil, make your edit, and press ⌘S. Naseem keeps track of unsaved changes for you, and if the file changes on disk while you’re editing, it asks before anything gets overwritten.

Enhanced in 1.21.0

• Browser testing just works. Screenshots from Chrome DevTools now save straight into your Outputs or project folder, first time.
• Test servers start fast and stay up. When Naseem spins up a local server to check your page, it gets going in seconds and keeps it running while it tests.
• Fewer interruptions. Naseem won’t ask you to approve changes to files it just created itself.

Smoother simulator work and easier MCP setup.

Better in 1.20.1

• Simulator apps launch right away. Apps on the iOS Simulator start at once, and Naseem says immediately if an app isn’t installed yet.
• Paste any MCP config. Add MCP servers by pasting the config from Claude, Cursor, VS Code or Codex — one server or several.
• Switch off built-in tools. Turn off Naseem’s built-in file or terminal tools in Settings → Agent, for example to use an MCP server instead.
• The MCP server editor responds to the first click.

Longer chats, fewer interruptions, and clearer costs.

New in 1.20.0

• Long chats & cache. Naseem now lets a chat you’re working in grow to about 250K tokens before summarizing its older part, because your provider already keeps a chat you’re using in its cache. Come back to a long chat after a break and Naseem summarizes it in the background while answering you. Choose a strategy in Settings → Agent → Long chats & cache: Auto, Keep more detail, Save cost, Custom or Off. Hover over the token counter in any chat to see the rule in force.
• Usage shows what the cache saved. The by-day chart now splits each day’s tokens into From cache, New and Not reported, and the Tokens in card shows the share served from cache.
• Costs are back, for Claude. Estimated costs are shown again, counting cached tokens at their real, lower price. For now they’re shown for Claude (Anthropic) only, whose prices Naseem has checked against a real bill; other providers show “—” until theirs are verified.

Better in 1.20.0

• Fewer “I’ve paused” stops. When a task repeats itself or hits errors, Naseem first tells the model to change approach. It only pauses and asks you if that doesn’t help.
• Simulator problems explained. Launching an app on a simulator that isn’t booted now says so at once, instead of waiting; a slow simulator driver tells the model what to check.
• The token counter after a relaunch shows the chat’s real size, or an estimate marked with “~” until the first reply.

Faster replies in long chats, and long tasks that keep going.

Better in 1.19.0

• Faster, lighter replies. Naseem now keeps each conversation steady from one message to the next, so Claude, ChatGPT and Ollama Cloud can reuse what they’ve already read instead of starting over. Replies start sooner and use less of your plan or budget, most of all in long chats and long tasks.
• Your edits apply right away. Changes you make in Knowledge (memory, lessons, skills, Naseem’s profile) reach your open chats on their next message. What Naseem learns on its own is saved at once and joins the conversation at its next summary or in your next chat.
• Long tasks keep going. Naseem no longer pauses a task as “not making progress” while it rebuilds and re-checks after a fix.
• Knows when to ask you. When something can’t be checked from the computer, like whether a sound played, Naseem says so and asks you instead of trying again and again.

Fixed in 1.19.0

• A finished reply no longer flashes twice for a moment.
• Naseem no longer creates a stray file when its answer mentions an Xcode project.
• The files list under a reply shows only the files Naseem worked on, not build leftovers.
• An edit that touches two files at once is split and applied cleanly instead of failing.

Smoother everyday use.

Fixed in 1.18.2

• Quit always works. Naseem quits even when the Knowledge window or another panel is open, so updates install without getting stuck.
• Notes save again. Naseem no longer refuses to update a project’s notes in some folders because it mistook the folder’s name for a password.
• Turn on a new Rule right in the chat. When Naseem sets up a Rule for you, a card under its reply shows the Rule with a Turn on button — one click and it runs on schedule.

A small fix to keep things running smoothly.

Fixed in 1.18.1

• Naseem no longer closes when a task needs something in Naseem itself. For example, while setting up a daily Rule, Naseem tried to approve the Rule on its own by clicking inside its own window, and closed instead. Now it asks you to approve the Rule yourself and carries on.
• Tab works like before in the message box. It no longer types spaces. When Naseem suggests what to ask next, Tab still fills it in, and Shift+Return still starts a new line.

A friendlier start, and memory you can trust.

New in 1.18.0

• Two modes: Everyday and Developer. Pick how you use Naseem; project chats follow their folder automatically. Switch any time from the sidebar.
• Starter cards. Six ready tasks above every new chat: summarise a document, turn receipts into a spreadsheet, explain a project, run the tests…
• A short setup when you first open Naseem. Choose your mode, connect a model, tell Naseem about yourself, pick a project and a language. Run it again any time from Help → Set Up Naseem…
• Memory that stays tidy. Naseem files what it learns after each reply: about you, your rules for Naseem, facts, or project notes. Everything shows in Recently filed, with Undo.
• Each project keeps its own memory. A project’s facts and lessons now live in its folder, in naseem/memory.md and naseem/lessons.md, next to its other notes. They travel with the project, and you can open and edit them in any editor. Naseem reads them fresh on every message. Existing project memory moves over when you choose: one review, a backup first.
• Edit Naseem’s memory directly. Knowledge → Memory lets you edit who you are and Naseem’s name, tone and principles.
• Tidy up memory and lessons. One review cleans up older, messy memory: you approve each change, and a backup is made first.
• Lessons that matter. Naseem learns from real mistakes and keeps each lesson where it applies: everywhere, for one tool, or in one project’s folder. No more review queue.

Better in 1.18.0

• Ollama cloud models are marked with a cloud icon.
• Brief network hiccups are retried automatically instead of interrupting the task.
• Keep a project’s Naseem files out of git with one switch in the project’s menu.
• Larger, easier-to-read hints across the new screens.

Smoother on real projects.

Fixed in 1.17.2

• Web projects go smoother. Naseem can now edit config files that allow comments, like tsconfig.json and VS Code settings, instead of refusing them and working around it.
• No more reading the same file in circles. When a long file or log is too big to show at once, Naseem now knows exactly which part it hasn’t seen and fetches just that part.
• New apps come with the basics. Ask Naseem to build an app or a page and it includes what you’d expect — a title, a readable layout, clear labels — while still asking before adding features you didn’t request.
• Clearer times. A paused reply shows how long the work ran, instead of “0 ms”.
• Ollama cloud models recognized. Models named like qwen3-coder:480b-cloud now get their real context size and can run helpers in parallel.

Write on more than one line.

Fixed in 1.17.1

• Shift+Return adds a new line. Return sends your message; Shift+Return (or Option+Return) starts a new line, so you can write lists and longer prompts in the composer. The hint under the box shows both.

Long chats that stay sharp.

New in 1.17.0

• Work as long as you like in one chat. When a conversation gets long, Naseem summarizes the older messages for the model and keeps the recent ones word for word. Long tasks stay fast and affordable, and the model keeps track of what you asked.
• Nothing you wrote is lost. Your whole conversation stays on screen. A thin line marks where the summary ends; click View summary to read exactly what the model sees. Your own requests and exact file paths are kept as you wrote them.
• Do it whenever you want. Type /compact in any chat, or add what matters most: /compact keep the API design. It’s also in the ⋮ menu.
• Choose when it happens in Settings → Agent → Compact conversations at. Auto picks the right moment for each model.
• Clearer steps. Each step in a reply now shows what it worked on — the file it read, the command it ran — and how it went, like “run_shell swift test — exit 0”.

Fixed in 1.17.0

• Enhanced long-context summarization. Very long chats no longer quietly forget how they started: earlier messages are summarized instead of dropped. And if a model says a chat is too long, Naseem summarizes and tries again on its own instead of stopping with an error.

Naseem makes images.

New in 1.16.0

• Create and edit images (Pro). Ask for a logo, banner, illustration or any picture, in any words or language, and Naseem makes it. Ask for changes and it edits the image it made, or one you attach.
• Use what you already have. Signed in with ChatGPT? Images come from your subscription, with no key and no extra bill. Or use your OpenAI or Google key. Pick in Settings → Models → Image model, or per chat from the model menu.
• Images right in the chat. Each image appears under the answer with the model that made it, and is saved to your Outputs folder and Artifacts.

Fixed in 1.16.0

• More honest answers after reopening a chat. Naseem now remembers what each earlier reply actually did, so a model can’t claim it ran something it didn’t. If an answer names an image that doesn’t exist, Naseem says so.

Faster, lighter, steadier.

New in 1.15.4

• Conversations open in under a second, however long they are. The model menu used to be rebuilt every time a chat opened; now it’s built only when you click it.
• Replies are easier to read. Each answer starts with Naseem’s name, like your own messages start with “You”, and ends with Copy, Read aloud, the model that wrote it and how long it took, right where you finish reading.
• Long chats stay quick. Very long conversations show the newest messages first; one click, “Show earlier messages”, brings up the rest.
• Project roadmaps keep dates. Each task now records the day work started and the day it was finished, so “what got done this week” has an answer.

Fixed in 1.15.4

• The chat view no longer goes blank, freezes or jumps, including in long conversations and after sending a second prompt. It now follows the newest content the way the system intends.
• Less memory on long and image-heavy chats. Screenshots in the conversation are kept as small previews instead of full-size pictures.

Claude is back to work.

Fixed in 1.15.3

• Claude can get things done again. Claude models kept stopping halfway, unable to run commands or open files. They now work through your tasks the way they should.
• You can see what Naseem is doing. While a longer step runs, like building an app, the chat now shows exactly what’s running and for how long, instead of a vague “Thinking…”.
• Saved files open in the side panel. Files Naseem creates in your home folder now show their contents in the Artifacts panel instead of “couldn’t read file”.

The transcript stays put.

Fixed in 1.15.2

• The transcript no longer goes blank or freezes after a reply or a new message. The conversation view had a layout loop that could stall the app from the third message on; it now follows the newest content the way the system intends.

Changed in 1.15.2

• Naseem now requires macOS 15 or later. Macs on macOS 14 keep 1.15.1.

A small fix for fresh installs.

Fixed in 1.15.1

• The MCP settings tab shows Add Server when no server is set up yet. The button used to sit below the window on a fresh install.

Z.ai, and Claude that thinks.

New in 1.15.0

• Z.ai (GLM) is a provider. Paste a Z.ai key, pick your plan’s endpoint, and chat with GLM-5.3 and its siblings, with reasoning effort and vision.
• Claude reasons now. Claude 4.6 and later think before answering, as much as your effort level allows.
• The model menu fits on screen. One submenu per provider, and every set-up provider lists its models the first time you open it.

Fixed in 1.15.0

• Long Claude answers are no longer cut off at 4,096 tokens.
• A Z.ai balance error explains what to do.

Better in Arabic.

New in 1.14.1

• Naseem speaks Arabic throughout. Start cards, shortcuts, new chat titles and labels now follow your language.
• Clearer when your data is newer. The message says what to do, and the button downloads the latest Naseem directly.

Fixed in 1.14.1

• The Skills list no longer repeats each skill’s description.
• A number like “~28.6k” is no longer mistaken for a missing file.
• Embedding models no longer appear in the model menu.

Sub-agents, side by side.

New in 1.14.0

• Sub-agents now work at the same time. Independent tasks run in parallel instead of one after another.

Fixed in 1.14.0

• A more dependable reviewer. It no longer reports “didn’t answer” by mistake, and it sees the commands each step ran.
• OpenAI models keep their reasoning between tool steps, so long tasks stay on track.

New in 1.13.10

• Leaner messages. Naseem now sends only the tools a chat needs; Calendar, documents, Slack and automation tools load when you ask for them.
• Project chats start informed. Your project’s rules (AGENTS.md, CLAUDE.md or naseem/rules.md) and notes come loaded, so work starts right away instead of re-reading them.
• Your notes are safe. Naseem’s automatic roadmap and journal updates only add and move items — they never delete anything.
• Checks that really ran. If your project declares a gate command, Naseem runs it before calling the work done.
• Clearer changes. The Changes card shows what’s committed, and lists apart any files someone else changed while it worked.
• Asks before risky git commands. Stash, checkout, restore, clean and rebase now ask first, even in chats that run without asking.
• Skills bring their tools. A skill that names the tools it needs loads them with it.

Fixed in 1.13.10

• GPT-5 models with your OpenAI key. They now work with tools and your chosen effort; the model list shows only chat models, and GPT-5.6 appears as Sol, Terra and Luna.
• Fewer failed edits. A simpler edit tool and clearer error messages mean far fewer retries.
• No more endless thinking. A step that only thinks for 10 minutes is told to act.

New in 1.13.0

• Slack. Ask what happened in a channel overnight, find the thread about the pricing change, see who asked about the release date. Naseem reads Slack as you, so it sees the channels you are in and nothing more, and there is no bot to invite anywhere. Posting is written in the conversation first and asks you every time, even in a chat set to run without asking — everyone in the channel sees it and it cannot be taken back. How it works. (Pro)
• Each project keeps its own memory. What Naseem learns inside a project — a decision, a convention, the order you want things done in — is filed under that project and brought back when you open it. What it learns about you stays with you everywhere. The Memory tab shows what is filed where, and every line can be moved in either direction. Nothing moved when you updated. How it works.
• The same PDF is only ever read once. Two helpers working on the same document now share what the first one extracted, instead of each doing the work again.

Fixed in 1.13.0

• Stop stops. Pressing stop while Naseem was inside a long step used to wait for that step to finish before it took effect. It now takes effect the moment you press it. Quitting no longer waits either: it saves what was written so far and closes.
• The transcript holds still. On models that call a dozen tools at once, the conversation could flash blank and back again while the steps landed. It stays where you are now, and tells you how many tools are running.
• Reading a file no longer costs more than it should. Naseem used to fetch long files in very small pieces — cheap-looking, and in fact the most expensive way to read, because every step re-sends the whole conversation. Ordinary files now come back in one go.
• Your Python is judged by the Python that will run it. A file Naseem wrote could be rejected as damaged when it was perfectly good, because it was being checked against the version of Python macOS ships rather than the one Naseem actually uses. And a genuine mistake now says what is wrong instead of asking for the same thing again.

New in 1.12.0

• Documents are read once. Point Naseem at a folder of proposals, specifications or reports and it reads each file a single time, writing notes you can open yourself: what the document is, the facts that matter with the page or slide they came from, and the questions it leaves open. Everything afterwards works from those notes, so the same file is never read twice and a long PDF does not fill the conversation. Read it again tomorrow and it costs nothing.
• A model chosen for reading. Settings ▸ Models has a Reading model: pick something with a long context and a low price, or a local model to keep the document on your Mac entirely. Leave it alone and Naseem picks a sensible one for you.
• A skill for starting from documents. Give Naseem a folder and a mission and it reads everything, writes the project description and the first three steps, then stops and asks you the questions only you can answer.
• Skills are yours to choose. The skills Naseem ships with now start switched off, because a skill changes how Naseem works and not just what it knows. Turn on the ones that match your work; two are marked as recommended, and one button turns those on. Every run names the skill it followed under the reply, where one click turns it off again.
• A Privacy chapter, in the manual and in the app: what stays on your Mac, everything that ever leaves it, and why. The full policy is one click from Settings.
• Integrations fold away. Each one is a card showing its state at a glance, and a new button at the bottom sends the developer the integration you actually need.

Fixed in 1.12.0

• Naseem no longer treats documents like code. Asking it to read a folder of files used to pull in a code-review procedure, split the work between helpers and read the same document several times over. It now reads them properly, once.
• Helpers are no longer handed tools they cannot use. A helper working on a model that cannot see images is no longer given an image to look at, and a helper that fails says so plainly instead of reporting success.
• Documents come back even when the reading model thinks too long.

New in 1.11.0

• Your Mac's apps. Turn them on one at a time in Settings ▸ Integrations ▸ Apple apps, and then just ask: what's on today, what came in this morning, find that note, remind me tomorrow at nine. Naseem uses the apps already on your Mac, with every account they sync — Gmail and Outlook included — so there is no account to connect and no key to paste. Nothing is fetched from the internet by Naseem, and nothing leaves your machine. How it works. (Free)
• Reading never asks; changing always does. Listing your events or searching your notes changes nothing, so Naseem gets on with it. Creating an event, adding a reminder or writing a note asks first. Deleting asks every time, even in a conversation set to run without asking, and so does sending an email — which is why Naseem would rather open a draft in Mail for you to send.
• ElevenLabs voices by language. Choose English or Arabic and the voice list follows, with the stock voices that speak every language kept beside the ones native to it. Browse ElevenLabs' Voice Library from inside Naseem and add a voice to your account in a click. (Pro)
• Integrations, tidied. Each one is a card you can fold away, with its state on the header at a glance: what's on, what's connected, what has no key yet.

Fixed in 1.11.0

• Naseem says when something is switched off instead of quietly doing it another way. Asking for a reminder while Reminders is off no longer puts an appointment in your calendar.
• The Voice chapter appears in the manual. It shipped in 1.10.1 but never showed up in the list.

New in 1.10.1

• Send by pausing. With Send after a pause on (Settings ▸ Voice ▸ Listening), going quiet for a few seconds sends your dictation — with a countdown under the composer for the last second. Saying Naseem still works, and is now its own switch too.
• A Voice chapter in the manual, covering speaking, dictating, the two ways to send, and voices — read it here.

Fixed in 1.10.1

• Naseem no longer hears itself. The microphone rests while Naseem answers and speaks, then comes back on its own for your next message — also when you stop the voice early.
• Settings search now finds the voice options by words like "pause" and "reply aloud".

New in 1.10.0

• Naseem speaks. Every reply has a speaker button, and Settings ▸ Voice has a new Reply aloud switch: turn it on — or click the speaker icon in the toolbar — and Naseem reads each reply as it finishes, in the reply's own language. Choose whether it reads the opening or the whole thing. Code and tables are never read out.
• Talk instead of typing. Click the microphone and speak. It keeps listening until you click it again, and what it hears lands in the composer for you to read. End with the word Naseem and the message sends on its own — a whole conversation with your hands free. Your voice is recognised on your Mac when the language allows it, and Naseem tells you when it isn't.
• Hear when a run finishes. With Speak a summary when Naseem isn't in front on, a run that ends while you're in another app tells you what it did — one sentence, never over you while you read.
• ElevenLabs voices, with your own key. Switch the engine in Settings ▸ Voice, paste your ElevenLabs key, and pick from their voices or your own cloned ones. The key stays in your Keychain; every sentence is billed to your ElevenLabs account, never to Naseem. (Pro)
• Settings, reorganised. A sidebar of eight panes with one search box that finds any setting by name. The Pro status sits in the footer, always in view.
• Knowledge inspector: bigger, larger text, and MCP tools grouped by server.

Fixed in 1.10.0

• The microphone now asks for permission. Earlier versions never showed the macOS prompt, so Naseem could not be allowed to hear you.
• Key fields in Settings are bordered and read left to right, with the cursor where you expect it.

New in 1.9.0

• Attach Word, PowerPoint and Excel files. Drop a .docx, .pptx or .xlsx on the composer and ask about it, the same way you already can with a PDF. Slides and sheets are labelled, so answers say where they came from, and spreadsheets come back as tables you can read. Everything is read on your Mac — nothing to install, and the file never leaves your machine. (Pro)
• Naseem Pro in the app menu. Buying, or managing your licence, is now one click away — directly under About Naseem.

Fixed in 1.9.0

• The reply follows along as it's written. The conversation now stays with the newest text while Naseem is answering, with or without an attachment — no more scrolling down by hand to find the rest.
• If a file can't be attached, Naseem says why. Over the size limit, a format it doesn't take, or a Pro format on the Free plan: the reason now appears under the composer, instead of nothing happening.