ncode A coding harness where every run, agent and operation is a process, and everything has a progress bar.

Ask for a change and watch it happen. Agents split the work, every tool call runs on its own, and nothing writes to your project beyond the approval mode you picked. One engine, as a native Mac app and as a full-screen terminal session.

Download for Mac Install the CLI

Developer preview for macOS 15 or later on Apple silicon. Bring your own model key.

⋔ swarm review changes 1 of 4 in 16:15

◌Lead waiting for 24/30
✓TS removal done8:34
◒Docs accuracy checking app data21/30
◒Strategy fit weighing 2 plansquiet 1m
✗Build check turn limit30/30

TS removal Deleting ailogic_typescript/ is safe: nothing in lib/ or assets/ imports it.

A swarm reviewing changes: the Lead waits on its reviewers, one is done, one ran out of turns. Rows from the CLI's side panel.

One engine, two faces

The Mac app and the CLI run the same engine on the same local database, so a conversation started in one is waiting in the other. They take turns with that database: quit one before you open the other.

ncode for Mac

A native window that draws a run as it happens: the transcript on one side, the agents working on the other.

ncode

ailogic

swarm review changes

let's plan how to make this app better

Remove the TypeScript client

Review the uncommitted changes before I commit.

⋔ swarm review changes running

One review is in. Build check ran out of turns, so the report will miss the build.

Run a command?

mix test

ApproveDenyDeny & stopAlways allow “mix test”

CardsTimelineChanges

Leadwaiting for 2

✓ TS removaldone

◒ Docs accuracychecking app data

◒ Strategy fitweighing 2 plans

✗ Build checkturn limit

The agents pane shows every agent as a card with its own bar, a timeline of what each one did, and the files the run changed.
Approve, Deny, Deny & stop or Always allow, answered right in the transcript.
A side chat beside any run (⌘⇧S) and a file finder over the project (⌘P).
Close the window and runs keep going; the menu bar icon brings it back. Scheduled tasks fire from the app.

ncode in the terminal

A full-screen session with the same runs, the same agents and the same approvals, kept to the keyboard.

● ailogic   ⋔ swarm review changes ● 16:15

 ⋔ Lead                         thinking ▌
   ├ ◒ Docs accuracy  thinking    running a command
   ├ ✓ TS removal     done        Deleting ailogic_typescript/…
   ├ ◒ Strategy fit   thinking    wrapping up.
   ╰ ✗ Build check    turn limit  no answer after 30 turns

 found           1 of 4 in · no files changed
 ⋔ ▄▄▄▄▄▄▄▄ ▁▁▁▁▁▁▁▁ ▁▁▁▁▁▁▁▁ ▁▁▁▁▁▁▁▁

 ▌                                             
 ^F agents  ^N needs you  ^B panel

A side panel of agents sorted by what needs you (Ctrl+B), and a hint letter to jump into any one of them (Ctrl+F).
Approvals by one letter: y once, Y for this run, A always for this command family.
Keep typing while a run works: a message steers the running turn or waits in the queue.
ncode -p "run the tests and fix what fails" --json runs one turn without the full screen, for scripts.

Six ways to send a message

The composer carries a mode. It decides what the next message starts: a turn that edits, a read-only plan, a goal, a workflow or a judged plan. Shift+Tab flips between Build and Plan in the app, and a slash command that names a mode selects it.

let's plan how to make this app better

🛠 Buildauto

What happens when you send

Every run is a small tree of processes. Stop it and everything under it stops with it.

It thinks

The agent reads your message, the project's instructions (AGENTS.md and friends, once you trust the project) and the project's memory, and decides what to do.
Each tool call is its own operation

Reading a file, a search, a command, a web page: every call runs on its own with its own bar, and the agent thinks again when they return. Reads are always allowed; writes and commands follow your approval mode.
A checkpoint before every write

Right before an agent changes a file, ncode keeps a copy of it. /rewind puts your files back the way they were before an earlier turn.
Big work fans out

A Lead splits the task across sub-agents that run in parallel, each able to work in its own isolated copy of the project, then checks what came back and reports.

Built for work that takes a while

Type / in the composer. These are the same in both faces.

/swarm <task>: A Lead breaks the task down, hands the parts to sub-agents running in parallel, then verifies and reports.
/goal <text>: Sets a goal every agent keeps in mind, across runs, until it is achieved.
/plan <task>: Plans with read-only tools. Approve the plan and it becomes the prompt for the build.
/consensus: A planner and a judge take turns until the judge approves, against checks you choose: scope, edge cases, keeping changes minimal.
/workflow <name>: Runs a workflow: a deterministic program of phases, agents and panels. Every step it takes on the host is journaled, so a stopped run can resume. /create-workflow writes one with you.
/review: Reviews the uncommitted changes and reports problems.
/compact: Writes a summary the model carries forward: the goal, decisions, files touched, open threads. Nothing in the conversation is deleted.
/rewind: Restores files to how they were before an earlier turn.
Deep research: Runs in the background, from Fastest (one round of four agents) to Ultra (four rounds of ten), and ends in a written report.
Scheduled tasks: A chat, swarm or workflow once, daily, weekly, monthly or on a cron line. They fire from the Mac app.
MCP servers: Add stdio or HTTP servers and switch off the single tools you do not want an agent to see.

Your key, your Mac

Bring your own model

Add a key for Anthropic, or point ncode at any OpenAI-compatible endpoint: OpenAI, OpenRouter, DeepSeek, or a model on your own machine through Ollama or LM Studio. Pick a model for each role: the chat, sub-agents, the judge, research.

Anthropic
OpenAI
OpenRouter
DeepSeek
Ollama
LM Studio
any OpenAI-compatible URL

Local first

Conversations, runs, settings and keys live in one database on your Mac. There is no account to create. ncode talks to the providers, search engines and MCP servers you add.

You set how far it goes

A new project starts read-only. Trust it to allow edits. In auto, writes are allowed and safe commands run while other commands ask; a dangerous command asks even in full access. Commands run with a scrubbed environment by default.

Get ncode

Version 0.1.0 is a developer preview for macOS 15 or later on Apple silicon. Take either face, or both. The downloads open with the 0.1.0 release; the guides are ready now.

The Mac app Coming soon

ncode-0.1.0.dmg · coming soon

Needs: macOS 15 or later, Apple silicon
SHA-256: TBD published with the download

This preview is not notarized, so macOS asks you to confirm the first launch in System Settings. How to open it

The CLI Coming soon

curl -fsSL https://code.llmotions.com/install.sh | sh

Needs: macOS 15 or later, Apple silicon
Update: run the same line again

The installer goes live with the 0.1.0 release. Then open a project with cd your-project && ncode. Install guide

Either way, the first thing to do is add your key: Settings → Providers & models in the app (⌘+,), or ncode settings providers in the terminal.

Read the guides

Mac app guideInstall, the window, runs and agents, workflows, settings. CLI guideInstall, the session screen, keys, headless runs, settings. ReleasesWhat changed in each version, with checksums.

ncode A coding harness where every run, agent and operation is a process, and everything has a progress bar.

ncode for Mac

ncode in the terminal

It thinks

Each tool call is its own operation