Notes from the work

One email when a tutorial or migration write-up goes live. Nothing else, and an unsubscribe link in every email.

LeanZero

Two people in Romania doing Atlassian migrations, Forge apps and practical AI work for teams that would rather talk to the person doing the job. Most of what we learn ends up on this site.

Services

  • Atlassian migrations
  • Atlassian FastShift
  • Atlassian maintenance
  • Forge app development
  • AI development consultation

Topics

  • Jira
  • Jira Service Management
  • Confluence
  • Bitbucket
  • Atlassian Forge
  • Cloud Migration
  • Local AI
  • AI Coding
  • Claude
  • Certifications
  • Atlassian Team
  • All topics

Company

  • Blog
  • Tutorials
  • Contact

Community

  • Join Discord
  • Support this site

© 2026 LeanZero. All rights reserved.

Privacy policyTerms of service24/7 SupportTrust centerSecurityDPALegal notice
  1. Home
  2. Portfolio
  3. Elzi
Tap me

Mac app for Apple silicon

Elzi

Formerly Goose Swarm

Hi, I'm Elzi.An agent that runson your Macs.

I write and test code on the Macs you already have, and benchmark the models that do it. When a job needs a cloud model, I use a provider you connect with your own key.

Download for Mac Explore all 18
Developer ID signed and notarized by Apple.

Up to 3.15×Up to 3.15×

Faster file edits when Elzi reuses text she has read.

Our 27B fine-tune on an M3 Ultra Mac Studio.

211211

Published benchmark runs, local and cloud.

On our public board at leanzero.net/agentic-benchmarks.

105105

Releases in 54 days.

Tagged from 19 September to 12 November 2026.

Eight things Elzi does, filmed in the real app.

Recorded on our MacBook and two Mac Studios on 11 November 2026, with music and no voice-over.

These are real recordings of the Elzi app on our Macs. Qwen3.8-Flash-Next runs in Elzi's own engine on our MacBook, with prompt lookup and the n-gram table on disk switched on. It answers a question about our benchmark data by running its own commands. A cloud model is added with our own OpenRouter key. One local model and three cloud models then share a swarm pool. Elzi cuts one brief into three parts, and two models write their own files at the same time. Claude, as the judge, sees a small model on a Mac Studio looping and sends it one next step. Elzi builds episode 5 of our YouTube series on Qwen3.8-Flash-Next while Claude Code supervises, both windows side by side. Claude Code's nudge lands in her chat and she makes the edit it asked for. A run of our Gauntlet benchmark with Qwen3.8-Flash-Next is live on the leanzero.net board, scoring 0.6932 in the same ranking as cloud models. The last two steps show Elzi's memories, which she keeps between chats, and her skills.

Pick a cell. Elzi flies you there.

Seventeen things Elzi does today and one on the way.

  • Thank you
  • Nodes
  • MLX engine
  • Prompt lookup
  • Memory lock
  • Agent Work
  • Swarm
  • The judge
  • Supervisor
  • LeanZero Link
  • Two Macs
  • Permissions
  • Tours
  • Benchmarks
  • Your models
  • Fast fixes
  • Images
  • Voice
  • Mobile remoteSoon

Devices1 of 18

Hand work to your Macs or a cloud model.

A node is a model plus where it runs: one of your Macs or a cloud provider. Connect any of 14 providers with your own key, then make nodes under Nodes.

Next: MLX engineNext
Elzi's Nodes view: a node is a model you can hand work to, on your Macs or in the cloud; New chats start on Qwen3.5-9B-MLX-4bit, this Mac (node)A node card in Elzi's Nodes view: Qwen3.5-9B-MLX-4bit on this Mac, MLX, Serving, idle; 39.2 tok/s writing, median of 104 runs, and 62.7 tok/s reading, median of 243 prompts
Elzi's Cloud Providers tab: configured, Gemini connected with gemini-4-argon and OpenRouter connected with deepseek/deepseek-v4.1-flash, each with a Change button; available to set up, OpenAI-compatible, marked Chat only: any server that speaks the OpenAI API; add as many as you need; make a node from any configured provider under Nodes; an Add endpoint buttonElzi's Cloud Providers tab, configured: Gemini connected with gemini-4-argon, and OpenRouter connected with deepseek/deepseek-v4.1-flash
Elzi's Nodes view, and two cloud providers connected with our own keys.

Engine2 of 18

Guessing ahead makes her writing up to 2× faster.

Elzi runs our own MLX engine on your Mac, a fork of open-source Rapid-MLX. Multi-token prediction lets a model built for it propose several word pieces and confirm them in one step.

Next: Prompt lookupNext

Same model, same Mac: writing speed with multi-token prediction

Code edit21.4 tok/s 21.4 without · 2.0×
Free chat22.0 tok/s 22.0 without · 1.7×
Summary with quotes22.1 tok/s 22.1 without · 1.5×

without multi-token prediction with it

Our 8-bit Qwen3.8-27B fine-tune on an M3 Ultra Mac Studio.
The Engine tab's Run it card: Run on this Mac, Best, Running; about 38.8 tok/s writing, 32.1 to 45.6 in the middle half, measured over 2 runs; 524,288 context; Stop and Measure speedThe Engine tab's Run it card: Run on this Mac, Best, Running; about 38.8 tok/s writing, 32.1 to 45.6 in the middle half; Stop and Measure speed

Engine3 of 18

File edits finish up to three times faster.

When Elzi rewrites a file she has read, the engine drafts the next stretch from it. The model checks the draft in one step and keeps only what it would have written.

Next: Memory lockNext

Change one number, return the whole 2,011-token file

Multi-token prediction alone 42.8 tok/s47.0 s0 tokens
With prompt lookup 134.8 tok/s14.9 s0 tokens

3.15× faster 1,926 of 2,011 tokens copied from the file and checked✓ the same output

With 8k to 256k tokens of other files read first

  1. 1.85×8k
  2. 1.74×32k
  3. 1.66×64k
  4. 1.44×128k
  5. 1.26×192k
  6. 1.24×256k

Fits bigger models104 → 72 GBQwen3.8‑Flash‑Next keeps its n‑gram table on disk, at nearly the same speed

Our 27B fine-tune on an M3 Ultra; long prompts and memory on Qwen3.8‑Flash‑Next, M5 Max.
Prompt lookup (n-gram) in Elzi's Sampling tab: Auto selected, with On and Off beside it, and the inside MTP badgePrompt lookup (n-gram) in Elzi's Sampling tab: Auto selected, with On and Off beside it, and the inside MTP badge

Engine4 of 18

Memory kept for your model while other apps run.

Turn on Memory lock under Providers > LeanZero MLX > Memory. A long chat then starts answering about as soon as on a quiet Mac, even while another app fills memory.

Next: Agent WorkNext

First word of a 145k-token chat while memory is tight

Memory lock on103.9 s+2.2 s from the other app
Memory lock off131.2 s+27.6 s from the other app

same turn on a quiet Mac added by the other app

Qwen3.5-9B on a MacBook Pro M4 Max, while another app held memory.
Elzi's Memory lock menu, open on Off, with Test and lock from the Engine tab, and Lock when the model mounts

Agents5 of 18

Agents that do their rounds without you.

Give an agent a schedule and press Start schedule. By default it wakes every 30 minutes on weekdays, does its round, and waits for you only when a decision is yours.

Next: SwarmNext
Elzi's Create an agent dialog: Purpose and workspace, where its research, drafts and history stay in its own folder, with a directory, a name, a title and a charter for who it speaks as, what it may touch and what it must ask about first; then Schedule: how often it wakes, set here to 30m (15m, 1h, 4h, daily or a custom interval also offered), a timezone, from 09:00 to 18:00, Monday to Friday selected, and an always open optionThe Schedule part of Elzi's Create an agent dialog: how often it wakes, set to 30m, with 15m, 1h, 4h, daily and custom to pick from; a timezone; Monday to Friday selected, Saturday, Sunday and always open not
Set in Elzi’s Create an agent dialog
When these settings wake itEvery 30 minutes, Monday to Friday09:0017:30MonTueWedThuFriSat, Sunno rounds
What one round leaves in its logSun 08:35 UTC tick 1 — Delivered the read-only report from lane charter-mdn-js-read without rewriting it.In plain words: a one-off Sunday test round read one web page and handed in its report as written.

The week is drawn from the dialog’s own settings.

Agents6 of 18

Three of our Macs built one app together.

In a September run, Elzi's swarm, a team of models, split one brief into 10 tasks across three Macs. All 10 finished: a payments console that draws 12,290 payments in 3D.

Next: The judgeNext
Each Mac's share of the run's 212 model calls, with tokens in and outInOutCalls
Mac Studio, M3 Ultra5.8M312k118
MacBook Pro, M4 Max3.0M117k51
MacBook Pro, M3 Max2.4M194k43
What the three Macs built: the Meridian Payments Console's 3D field of payments, thousands of small coloured blocks laid out in rows on a dark canvas, with a few of their amounts, in Kuwaiti dinar and dollars, floating over itWhat the three Macs built: the Meridian Payments Console, 12290 payments; 3201 payments, 2,175,738.05 euro; 3038 payments, 1,972,187 yen; 3023 payments, 2,043,135.720 Kuwaiti dinar; 3028 payments, 1,961,647.62 dollars

What they built: the Meridian Payments Console's 3D field of payments.What they built: the Meridian Payments Console, holding all 12,290 payments in four currencies.

Qwen3.8-27B on each Mac, from the run's public page.

Agents7 of 18

One model stands guard so others don't loop or drift.

In a swarm build, a guard reads what each working model is thinking. If one drifts or loops, the guard sends it one next step; it never stops the task itself.

Next: SupervisorNext
  1. VERDICT|OK|HIGH|Read all 873 lines...
  2. DRIFTING|HIGH|Answer is fully worked out in its reasoning: role matrix...
  3. One steer ('call the output tool now with the current task-DAG table containing only ids, files, and dependencies'), then 2 min later the plan: 7 tasks. The whole plan-writing call took 8.1 min.

Mac Studio M3 Ultrastanding guard

  • Mac Studio M3 Ultraits own build work
  • MacBook Pro M4 Maxworking
  • MacBook Pro M3 Maxworking

What the guard said, word for word

OK“Read all 873 lines...”a model planning the build
DRIFTING“Answer is fully worked out in its reasoning: role matrix...”a model researching one question
Steer

“call the output tool now with the current task-DAG table containing only ids, files, and dependencies”

to the model writing the build plan

In plain words: send the plan now, just the tasks, their files and what depends on what.

2 min later

the model did it and sent its plan: 7 tasks.

Quoted from our notes on two past swarm runs, on three of our Macs.

Agents8 of 18

Claude Code keeps watch.Elzi does the work.

Claude Code watches long Elzi jobs on the same Mac, and a job counts as done only when its checks pass. You open Claude Code from Supervisor in Elzi's sidebar.

Next: LeanZero LinkNext
An Elzi session with a strip on top: Supervised by Claude Code, Watching, job workmule-20261110-071923-84e2, The Overseer is connected and waits; the engine keeps watch and wakes it. Under it Elzi's own work: Thinking, then Shell, extracting episode line plan data, Completed; Thinking, then Shell, editing LINES.md with python, CompletedThe strip on a supervised Elzi session: Supervised by Claude Code, Watching

Claude Code is connected to this job and waits while Elzi works on it.

Devices9 of 18

Every Mac in the house,in one list in Elzi.

Sign in to LeanZero Link on each Mac with your email and a 6‑digit code. In My Macs you then choose which Mac runs a model.

Next: Two MacsNext
Elzi's My Macs tab seen from a MacBook: Your Macs, 3 connected; Models, 8 on your Macs; Run it, Qwen3.5-9B-MLX-4bit running; Nodes, 3 nodes; and the line Every Mac signed in to your LeanZero Link account, load models on any of them from the Models tab and pick where a model runs from the Engine tabStep 1 of Elzi's My Macs setup, done: Your Macs, 3 connected
  1. 64 GBThis Mac
  2. +96 GB
  3. +128 GB
64160288 GBacross 3 Macs, not pooled

From My Macs on our MacBook Pro M3 Max.

Engine10 of 18

Run a model too big for either of your Macs.

Join two Macs with a Thunderbolt 5 cable and Elzi splits one large model across them, part of its layers on each.

Next: PermissionsNext

One model, two Macs, drawn to scale

161 GBof weights: 33 GB more than the bigger Mac's memory.

engine limit
84 GB
17 layerspeak 66.8 GB
engine limit
118 GB
31 layerspeak 114.6 GB33 GB too bigMiMo-V2.6-Flash
M3 Ultra Mac Studio96 GBThunderbolt 5M5 Max Mac Studio128 GB

in memory at the run's startworking memory the run added, up to its peak

5 h 02 min of agent work145 model calls389,852 tokens in the longest prompt

Our build of Xiaomi's MiMo-V2.6-Flash on two Mac Studios, one five-hour agent run.
The result card of the two-Mac run in Elzi's Benchmark view: Gauntlet 7.2, payments, current, 4 November 3:21 PM; score 53.7 percent, model build 5 h 2 min; finished on its own after 145 of 150 model callsThe two-Mac run's result card in Elzi, stacked: Gauntlet 7.2, payments, current, Nov 4, 3:21 PM; score 53.7 percent; model build 5 h 2 min; finished on its own after 145 of 150 model calls
The run's result card in Elzi's Benchmark view.

Devices11 of 18

Each of Elzi's macOS permission requests says who is asking.

Elzi carries seven of these prompts. Three are for things Elzi does itself; four appear only when a program you run through Elzi asks.

Next: ToursNext
  1. Elzi 4.1.7 net.leanzero.agent
  2. NSAppleEventsUsageDescription: Elzi controls another app only when you ask, through an agent or in Free Memory.
  3. NSAudioCaptureUsageDescription: Elzi doesn't use your computer's audio itself. A program or app you run through Elzi is asking for it.
  4. NSBluetoothAlwaysUsageDescription: Elzi doesn't use Bluetooth itself. A program or app you run through Elzi is asking for it.
  5. NSCameraUsageDescription: Elzi doesn't use your camera itself. A program or app you run through Elzi is asking for it.
  6. NSLocalNetworkUsageDescription: Elzi reaches your other Macs over Thunderbolt and your local network to run distributed inference and to copy models between them.
  7. NSMicrophoneUsageDescription: Elzi listens only while you dictate into the chat box, or if a program or app you run through Elzi asks for the microphone.
  8. NSRemindersUsageDescription: Elzi doesn't use your reminders itself. A program or app you run through Elzi is asking for it.
3 Elzi uses it
  1. Local Network“Elzi reaches your other Macs over Thunderbolt and your local network to run distributed inference and to copy models between them.”
  2. Microphone“Elzi listens only while you dictate into the chat box, or if a program or app you run through Elzi asks for the microphone.”
  3. Automation“Elzi controls another app only when you ask, through an agent or in Free Memory.”
4 Only if a program you run through Elzi asks
  1. Camera“Elzi doesn't use your camera itself. A program or app you run through Elzi is asking for it.”
  2. Reminders“Elzi doesn't use your reminders itself. A program or app you run through Elzi is asking for it.”
  3. Bluetooth“Elzi doesn't use Bluetooth itself. A program or app you run through Elzi is asking for it.”
  4. System audio recording“Elzi doesn't use your computer's audio itself. A program or app you run through Elzi is asking for it.”

Quoted word for word from the signed Elzi app.

Proof12 of 18

Elzi shows you around on day one.

The first-run tour shows you what Elzi can do in 11 steps. Along the way you make a practice node that can't answer anything, and at the end she clears it away.

Next: BenchmarksNext
Step 6 of 11 of Elzi's first-run tour: That's your practice node. It's marked Practice and it can't answer anything, because its provider is made up. I'll remove it when the tour ends.
123introduce4567practice node8910the MLX engine11wrap-up
  1. 1
  2. 6
  3. 11

12menus have their own short tour the first time you open them.

From the first-run tour on our MacBook Pro M3 Max.

Proof13 of 18

Rank a model on your Mac against the cloud.

In Elzi's Benchmark view every model, local or cloud, gets the same frozen build task. What it builds is run and scored against fixed checks, never by another model.

See the boardNext: Your modelsNext

All 51 runs of the current Gauntlet, by rank

Ours: 0.6932, 15th of 51with wrap-up reminders, which 48 of the 51 runs didn’t get
rank 1rank 51

cloud run local run ours

Just behind Grok 4.7 and ahead of 32 of the 46 cloud runs. The top run scored 0.9574.

Our 4-bit Qwen3.8‑Flash‑Next on one M5 Max Mac Studio, from our public board.
Rows 14 to 16 of the Gauntlet 7.2 leaderboard on leanzero.net: Grok 4.7 at 0.6937, our local Qwen3.8-Flash-Next with wrap-up nudge at 0.6932, Gemini 3.8 Flash at 0.6929

Proof14 of 18

Put your fine‑tune through the same test.

Load your own MLX fine-tune into Elzi and run it through the Gauntlet, the build task frontier models are graded on. Press Publish to post its run to our public board.

Ask us about your runNext: Fast fixesNext
Elzi's Benchmark view, New run: Benchmark, Gauntlet 7.2 payments; Swarm chosen; Nodes 1Elzi's Benchmark view, New run: Gauntlet 7.2 payments, Swarm chosen, Nodes 1

2. In Benchmark, pick the Gauntlet, then Swarm and 1 node.

One row of Elzi's Models tab: our fine-tune Mihai-LeanZero/Qwen3.8-27B-Atlassian-v2, 31 GB, On diskOne row of Elzi's Models tab: our fine-tune Mihai-LeanZero/Qwen3.8-27B-Atlassian-v2, 31 GB, On disk

1. Your fine-tune, on disk in Elzi's Models tab.

Its row on the run's board: Mihai-LeanZero/Qwen3.8-27B-Atlassian-v2, the 8-bit MLX build, single agentIts row on the run's board: Mihai-LeanZero/Qwen3.8-27B-Atlassian-v2, the 8-bit MLX build, single agent

3. Our run's row on its board; yours reads Your fleet · 1 node.

  1. 1

    1. Your fine-tune, on disk in Elzi's Models tab.

  2. 2

    2. In Benchmark, pick the Gauntlet, then Swarm and 1 node.

  3. 3

    3. Our run's row on its board; yours reads Your fleet · 1 node.

Our 8-bit Qwen3.8-27B Atlassian fine-tune, on an M3 Ultra Mac Studio.

Proof15 of 18

A bug we find ships fixed within hours.

We timed seven bugs found over two days of releases: all are fixed, and six shipped in the very next release. Updates keep your chats, models and settings.

Next: ImagesNext
The Elzi release notes on leanzero.net: 4.1.7 and 4.1.6 on 7 November, 4.1.5 and 4.1.4 on 6 November 2026, each with what changedThe Elzi release notes on leanzero.net at phone width: 4.1.7 and 4.1.6, 7 November 2026, each with what changed

Half the bugs shipped fixed within 5.3 hours of being seen

first seen1 hshipped fixedmedian

  1. Chat said no node could answer while the engine served another modelfixed4.4 h
  2. The large app icons stuck in a grey tile on macOS 27fixed2.6 h
  3. The first-run tour kept coming back, also after an updatefixed5.3 h
  4. The model list stuck on “Reading the models folder”fixed5.3 h
  5. One slow node ended your chat turnfixed10.8 h
  6. After an update, Link stayed disconnected until you pressed Retryfixed10.6 h
  7. A chat whose node was removed showed the node's raw idfixed19.6 h
0 h5 h10 h15 h20 hfixed00.0 h

Elzi checks for updates every 4 hours and installs the new version the next time you quit.

Times from our notes, logs and git tags.

Media16 of 18

Elzi draws pictures on your own Mac.

Add the free mflux image tool, turn on Image Generation and ask in plain words. Elzi keeps each picture with its prompt and seed, and sends nothing out.

Next: VoiceNext

Seven cartoons, timed from request to picture on our Macs

Confluence Page Approval Workflow, Step by Step

  • Robots at an approval desk0 min 0 s
  • A robot repainting a picture0 min 0 s

Jira Automation vs Workflow Validators

  • A robot flipping a conveyor lever0 min 0 s
  • A robot clerk stamping a receipt0 min 0 s

Jira OKR Tracking Without an App: Atlassian Goals

  • A robot pointing at task cards0 min 0 s
  • A two-panel comic about a key0 min 0 s
  • A robot painting a progress bar0 min 0 s
One picture in Elzi's Images view: robots at a factory receipt board, with the start of its prompt, 'A cheerful comic book scene at a giant factory receipt board: a friendly', its date, 6 November 2026, and an Open session link
Each picture is kept in Elzi's Images view.

Media17 of 18

A voice of her own,made on our Macs.

She narrates our YouTube episodes in a voice made by an open speech model on our Macs. It's our own video tool, not part of the download.

Next: Mobile remoteNext

Picked by ear, not by score

  1. 15voices tried
  2. 4models in the test round
  3. 3blind samples, from 2 models
  4. 1kept, though 4th of 15 by score
Our audition: a 3‑minute script and a 20‑minute drift test.

Devices18 of 18Coming soon

Steer your Macs from your phone.

Our iPhone and iPad app will show every Elzi session on your Macs. You'll answer an agent, steer or stop a run, and the session keeps running on your Mac.

It's in testing and not on the App Store yet.

Thank you to the open source we build on.

Elzi is our own app. Three open-source projects gave us the most, and we're grateful to the people behind them.

  • Rapid-MLX

    By the Rapid-MLX contributors; it began as Wayner Barrios's vLLM-MLX · Apache-2.0

    The biggest piece we borrowed: Elzi's MLX engine is our fork of it.

  • goose

    Created by Block, now at the Agentic AI Foundation · Apache-2.0

    Where Elzi started, as Goose Swarm.

  • MTPLX

    By Youssof Altoukhi · Apache-2.0

    Her multi-token prediction uses code from it.

Elzi is based on goose and Rapid-MLX and uses code from MTPLX, all under the Apache License 2.0; not affiliated with or endorsed by the goose project, the Agentic AI Foundation, the Linux Foundation, Block, the Rapid-MLX authors or the MTPLX authors.

She has her own show.

Elzi presents our YouTube episodes, in the voice she makes on our Macs. Her name is LZ, LeanZero's initials said aloud.

Watch on YouTubeAll 5 episodes
How we made Elzi, our unpaid voxel AI YouTube narrator, on a local TTS model
ArticleLocal AISentinel Vault

How we made Elzi, our unpaid voxel AI YouTube narrator, on a local TTS model

Elzi is the little voxel robot who narrates our YouTube videos. She was born yesterday, she's still waiting for a salary, and her first video came down two hours after it went up. Here's how the two of us built her in five days on a local TTS model, what's still wrong with her, and why we want your ears.

Oct 21, 202632 min read
  • ElziAtlassian Champion Unboxing: Gabriela's Welcome Box | Elzi Bonus
  • CogniRunnerAn AI Coding Agent in Jira That Asks Before It Commits
  • CogniRunnerRestore Deleted Jira Tickets? Our Guard Dog Kept a Copy
  • Sentinel VaultConfluence Deleted Page Not in Trash? What Still Survives

Meet Elzi on your own Mac.

For Apple silicon Macs, signed with our Developer ID and notarized by Apple.

Download for Mac

Release notesFeature guides