1.21.3Native Swift · No Electron · macOS 15+

The Mac agent that does the work, not just the chat.

Naseem is a fast, lightweight AI agent for Mac — a gentle breeze against heavy, hot Electron assistants. Chat with any major model or fully on-device, and let it safely act on your Mac.

Download for macOS Get Naseem Pro $49

Free planNo accountNo subscriptionBring your own AI key

Naseem is a Mac app (macOS 15+). On Windows or a phone? Follow Naseem on Product Hunt for updates, or send yourself the link for later.

weather-cli · Software Engineer Autonomous

In ~/NaseemDemo/weather-cli: the temperatures in FORECASTS are Celsius but the CLI prints them as °F. Fix the unit, add a --fahrenheit flag that converts, have a sub-agent write tests for both units in test_weather.py, then run the tests. Don't commit.

make_plan6 steps, each with a check

read_fileweather.py, test_weather.py

edit_fileweather.py · adds --fahrenheit

delegate_taskwrite the unit tests

run_shellpython3 -m unittest → 5 OK

Tests: all 5 passed. No git commands were run — changes are left in the working tree for you to review.

Review · passed

A real run, not a mockup — see it below ↓

An agent, not just a chat window.

A powerful model alone isn't an agent — it needs a harness. Naseem is that harness, plus the one thing browser-bound assistants don't have.

LLMthe intelligence

Agent harnessthe agency

Your Macthe power to act

See it work

Real runs, not mockups.

Every picture below is from a real Naseem chat on an ordinary model, captured as it happened. Click one to see the whole chat.

Naseem’s answer: what changed in weather.py, the four tests a sub-agent added, all 5 tests passed, and the review passed — **Plan, work, prove it**It fixed the bug, had a sub-agent write the tests, ran them — 5 passed — and a reviewer checked the work before it said done.

The Changes card with the red and green lines of the weather.py diff — **Every change, line by line**The Changes card shows exactly what was edited, file by file. Nothing is committed — that stays yours.

Asked to commit, Naseem shows the exact git command and waits for Allow or Don’t allow — **Git history asks first**Even when a chat runs tools on its own, commits, pushes and resets always wait for your yes.

Sub-agents at work, each on its own research task, with its model and tool calls shown — **Sub-agents**Independent pieces of a task run side by side, each in its own sub-agent, with its steps visible as it works.

Naseem next to the iOS Simulator, having opened Settings, General, About and read the iOS version — **Drives the iOS Simulator**Asked for the iOS version, it opened Settings → General → About on an iPhone 17 Pro and read it off the screen.

Naseem next to Calculator showing 1,234 × 5,678 = 7,006,652 — **Uses your Mac apps**It pressed Calculator’s own buttons to work out 1,234 × 5,678 — then checked the display before answering.

Naseem in Arabic, laid out right to left: asked in Arabic for a small Python script, it wrote and ran it and answered in Arabic that the average of 3, 7 and 11 is 7.0 — **Speaks your language**The whole app in Arabic, laid out right to left. Asked in Arabic, it wrote the script, ran it, and answered in Arabic.

Why Naseem is different

Most assistants stop at an answer. Naseem keeps going.

Inspect the project, make the change, run the tools, read the results, verify, and fix — on your real files, in your real terminal.

Works with your real files

Inspect, read, patch and create files on disk. No copy-pasting between a chat window and your editor.

Has a real terminal

git, xcodebuild, package managers, tests, scripts, Python. Your Mac's actual environment, not a browser sandbox.

Drives the Simulator

It sees the app's real buttons and labels and acts on them by name — not by guessing from screenshots.

Drives native Mac apps

The same approach through Apple's Accessibility, for GUI-only apps like Notes, Mail and System Settings.

Proves, doesn't assert

A "done" is a claim to check. A file that isn't on disk sends the reply back; a tap on a changed screen is refused.

Stays fast on long tasks

Hundreds of steps without the conversation getting heavier and slower with every one.

Nothing claimed early

Feedback from the forum has become releases within hours. A principle the app doesn't keep counts as a bug.

Conductor, not orchestra

Written entirely in Swift. It stays light while Xcode, the Simulator and the model do the heavy lifting.

Long-task engineering

Built to stay sharp over hundreds of steps.

The hard part of an agent isn't answering — it's staying useful without drowning in its own context or losing the plot.

Light on every step

Finished work doesn't pile up, so a long run stays quick and affordable.

Memory that survives restarts

Ask about yesterday's failed build and the agent pulls up the actual log.

Stays on mission

It keeps the goal in view on every step, and stops to report back instead of grinding.

Crash-safe runs

Quit mid-task and completed steps are kept; continue picks up with memory of what was done.

Naseem Settings, Models: How Naseem works — Planner, Main model, Sub-agents, Reviewer — Planner→Main model→Sub-agents→Reviewer
Give each role its own model: a strong one to plan and review, a cheaper one for the bulk of the work.

Everything in the box

Free is the agent you supervise.
Pro is the agent you hand work to.

FreePro

iOS Simulator driving

PRO

Boot, build, install, tap and swipe through the real UI, read logs, take screenshots.

Mac computer use

PRO

Drive GUI-only Mac apps via Accessibility — allowlisted per chat, approved per action.

Sub-agent delegation

PRO

Hand focused pieces of a task to child agents that work side by side, with their steps shown live.

Images

PRO

Logos, banners, illustrations — made and edited in the chat, with your ChatGPT subscription or your OpenAI or Google key.

Telegram remote

PRO

Talk to your agent from your phone, including remote Approve / Deny.

Rules

PRO · BETA

Scheduled runs that stay quiet until something needs you — and never say "all clear" without running the check.

Profiles

Assistant, Writer, Software Engineer, QA, Automation — each with its own tools and approval policy.

On-device inference

Fully local Apple-Silicon inference with MLX. No network, private by construction.

Native tools

Terminal, files with diff-based patching, Python, PDF — behind per-call approval. Rarely used tools load only when a chat needs them.

Arabic, right to left

Switch the app to Arabic and everything follows: the layout mirrors, the start cards and labels are translated, and answers come in the language you write.

Every provider

Anthropic, OpenAI, Gemini, Z.ai (GLM), OpenRouter, Ollama Cloud, local Ollama, on-device MLX.

MCP

Connect any Model Context Protocol server with a paste-whole-JSON config editor.

Memory & skills

Cross-conversation memory; recurring tasks become reusable skills you review.

Web search

Built-in web search for news, docs and prices — bring a free Tavily or Brave key.

Autonomy per chat

Ask-first or unattended, set per conversation. Profiles can lock it down, never loosen it.

Vision for any model

A text-only model still inspects screenshots through your chosen vision model.

Start cards & notes

New chats open with tasks drawn from the project notes Naseem keeps for you.

Usage & Cost

BETA

Measured tokens per call, live prices and a memory-footprint panel.

Naseem updates itself, and every release is documented. See what's new.

Pricing

Free to start. Pro if it earns it.

Naseem uses your own models and API keys, so buying Pro doesn't lock you into another monthly AI bill.

Free

$0no card

Every provider, including local MLX
Conversations & up to 3 projects
Web search, vision, artifacts
Core tools with approval, MCP
Memory & skills, Quick Ask

Download Free

Pro

$49lifetime

iOS Simulator automation
Native Mac computer use
Sub-agent delegation
Image generation & editing
Telegram remote access
Self-improving skills & custom profiles
Unlimited projects · up to 3 Macs

Get Naseem Pro 14-day Pro trial · No credit card · No subscription