MimiWork

The AI coworker that hands you the finished file.

Not a chat window. A colleague who reads your folder, does the work, and saves a .docx, .pptx or .xlsx you can send — on your machine, with your model key, asking before anything consequential.

---

The difference in one line

"Summarise these three PDFs into a one-page brief."

Most assistants: a summary in the chat, which you then paste into Word and format. MimiWork: brief.docx in your folder — PDFs actually read, tables extracted, numbers carried across, formatting intact.

That's the whole idea. Ask for an outcome, get the artifact.

Why it feels different

Actually agentic. Not a chat window with tools bolted on: Mimi plans the work, reads the files, runs the analysis, checks its own layout, and iterates until the file is finished — dozens of steps from one ask. Plan mode shows you the whole approach before anything runs.

It learns as you work. Memory keeps what it noticed about your projects. Skills package your know-how once — brand rules, methods, house style — and apply forever. Instructions set standing rules per folder. Week two is visibly better than week one, and everything it learned is yours to read, edit, or delete.

Steerable while it works. See it heading the wrong way? Just type. Your correction lands at the next safe step without restarting the work — and a looping model is stopped before it burns your credits.

It runs while you don't. Say "every Monday at 8…" in plain words and it becomes a local automation with a full transcript; anything needing a decision waits in your Inbox instead of guessing.

Yours, on your machine. No vendor cloud, your model keys, every consequential action approval-gated, and nothing outside the folders you grant is readable.

What it actually does

One coworker, three trades. There is no mode to pick — Mimi carries all of this into every conversation and reaches for the right craft when the work calls for it.

	What it produces
Data analysis	Profiles the dataset before touching it — SPSS/Stata variable and value labels included, so q4_1 becomes "Satisfaction with onboarding, 1–5" — runs the analysis in a persistent Python kernel or R, and reports sample sizes, effect sizes and assumption checks, not just p-values.
Documents	Drafts and revises Word documents in place without wrecking your formatting — headings, tables, comments and all — and can leave real tracked changes with a plain-language "what I changed and why".
Decks	Turns findings into a 16:9 PowerPoint that argues a case — every slide titled with its takeaway, not its topic, and built from a real layout vocabulary (a single claim, a big number, a participant's own words, a side-by-side, a chart) rather than slide after slide of bullets. Charts come from your actual numbers, and speaker notes come with it. Text is measured and fitted so nothing walks out of its frame. Point it at your house template and it uses your brand instead.

It reads what work actually arrives in. PDF (with table extraction — and an honest warning when a document is scanned rather than silently returning nothing), Word, Excel, PowerPoint, CSV, SPSS .sav, Stata .dta, and images it can crop, annotate and combine into figures.

Teach it your way once. Package your brand colors, fonts and house rules into a skill and every deck, document and spreadsheet comes out in them without being asked. Browse 8,400 community skills by shelf, and read a skill's real instructions — and which tools it wants — before it ever lands on your disk.

It connects to where you work. 25+ integrations — Slack, Gmail, Outlook, Google Calendar and Drive, GitHub, Jira, Notion, Linear, HubSpot, Canva, monday.com — plus anything reachable over MCP. Tag it in a Slack channel and the finished work comes back as a thread reply. Canva signs in with one click and Mimi can search, generate, edit, comment on and export your designs there.

Including where your data lives. Connect Qualtrics and ask for the December wave: it reads the questionnaire so Q4_1 arrives as "How satisfied were you with — speed of setup", then downloads the responses as a CSV or a labelled SPSS .sav in your folder, ready for the analyst. It never edits or sends anything in Qualtrics, and the download itself asks you first.

It runs while you don't. A Monday brief, a weekly report, a standing watch on a channel: automations run locally with full transcripts, and anything needing a decision waits in an Inbox instead of guessing.

You can steer it mid-run. Notice it heading the wrong way? Just type — "use the December wave, not November" lands at the next safe step without stopping the work, and a looping model is caught and stopped before it burns your credits.

It checks in before acting. Sends, writes, shell commands and external data pulls are approval-gated. Plan mode proposes first and runs nothing until you say go.

What you learn here works everywhere

MimiWork deliberately borrows the vocabulary and the gestures of Claude Code, Claude Cowork and Codex — so the habits transfer in both directions, and nothing you learn is trapped in one app.

In MimiWork	Cowork	Claude Code	Codex
/ — app commands, your saved commands, your skills	/ commands	slash commands	/ commands
@ — point at a file in a granted folder	@ mentions	@ mentions	@ mentions
Plan · Ask for approval · Full access (⇧⇥ cycles)	Manual / Auto / Skip	plan / default / bypass	plan / approval / auto
AGENTS.md — and CLAUDE.md is read too	folder instructions	CLAUDE.md	AGENTS.md
Skills — SKILL.md folders, importable from ~/.claude/skills	Skills	Skills	—
Projects — folder + instructions + memory + threads	Projects	a repo you cd into	a repo you cd into

Settings ▸ Transfer guide keeps the full map one click away.

Bring your own model — no lock-in, no markup

Paste your own key and switch anytime:

OpenAI · Anthropic · Google Gemini · DeepSeek · Qwen · Kimi (Moonshot) · GLM (Z.ai) · MiniMax · Mistral · Grok (xAI) · Inkling — open-weight models via Together and Fireworks, AWS Bedrock, and fully local via Ollama.

Or skip keys entirely: sign in with your QualiTaTi account and the "Mimi" models spend your existing credits — nothing to configure. The Activity page shows exactly what each call cost and which pool paid — the numbers come from the server's own ledger, not a local estimate. A switch in Settings picks where the models run: US (cheaper credits) or strict-GDPR Paris (data stays in Europe). Signing in also opens your QualiTaTi research data: ask for a project, an interview transcript or a survey's responses and Mimi can pull them in to analyse — each retrieval asks your approval first, and nothing is fetched on its own.

Install

The app updates itself from then on.

Your machine	Download
Mac — Apple Silicon (M1–M4)	MimiWork-macos-arm64.dmg
Mac — Intel	MimiWork-macos-x64.dmg
Windows 10/11	MimiWork-windows-setup.exe

MimiWork is not yet signed with Apple or Microsoft, so both systems warn about it the first time. Nothing is wrong with the file — the warning only means the publisher is unknown to them. Here is what you will see and what to click, on current macOS (Sequoia 15 and Tahoe 26) and Windows 11. It is a one-time thing: updates install from inside the app and never trigger it again.

Mac — the first launch

1. Open the .dmg and drag MimiWork into Applications.
2. Double-click MimiWork. macOS says “MimiWork” Not Opened — Apple could not verify… or “MimiWork” is damaged and can’t be opened. Click Done (never Move to Trash).
3. Open System Settings ▸ Privacy & Security and scroll down to Security. Under “MimiWork” was blocked to protect your Mac, click Open Anyway. The button stays for about an hour after step 2; if it is not there, double-click MimiWork once more and come back.
4. Double-click MimiWork again, click Open Anyway in the dialog and enter your Mac password. macOS remembers the choice.

If the Security section shows nothing, or the Mac is managed by your organisation: open Terminal (⌘ Space, type Terminal) and paste

xattr -cr /Applications/MimiWork.app

then open MimiWork normally. That removes the “downloaded from the internet” flag the browser put on the file, which is all the “damaged” message is about.

Windows — the first launch

1. Download MimiWork-windows-setup.exe. The browser may hold the download back:
   • Edge: click the download at the top right, then ⋯ ▸ Keep ▸ Show more ▸ Keep anyway.
   • Chrome: click the download arrow, then Keep (sometimes behind Show more).
2. Run the installer. Windows shows the blue Windows protected your PC screen. Click More info, check that the app is MimiWork-windows-setup.exe, then Run anyway.
3. Finish the installer. MimiWork opens.

If Run anyway is missing (a work PC where SmartScreen is locked by policy): right-click the downloaded file ▸ Properties ▸ General ▸ tick Unblock ▸ OK, then run it. If that box is missing too, your IT team has to allow it (a Microsoft Defender SmartScreen exception or an allow-list entry for MimiWork).

Apple’s and Microsoft’s publisher certificates are paid, yearly things; the release pipeline already signs and notarizes the moment QualiTaTi has them, and these steps go away. Every release, with checksums and older versions, is on the Releases page.

The ten-minute tutorial

The one habit that matters most: ask for the outcome, not the steps. "Read these transcripts and write me a themed summary as a Word doc" gets you summary.docx. "Can you help me analyse interviews?" gets you a conversation. (One page: docs/TUTORIAL.md — also in 中文, Norsk, Français · Illustrated PDF, 13 pages: docs/MimiWork-Tutorial.pdf)

1 · Connect a model — QualiTaTi credits or your own key (2 minutes)

Open Settings ▸ Models. Two ways in:

• Sign in with QualiTaTi — no keys; the Mimi models spend your existing credits. The card shows the three tiers — Mimi Puppy (free every day), Mimi Hound (fast, also free), Mimi Wolf (most capable) — each with a Test button that makes a real call. While you're here: pick your model region (Default · US, cheaper — or Strict GDPR · Paris, data stays in Europe), and know that the Activity page shows exactly what each call cost, from the server's own ledger.
• Paste your own key — OpenAI, Anthropic, Gemini, Kimi, DeepSeek, Mistral and more, or fully local via Ollama. Switch anytime from the composer.

2 · Give it a folder — the one real setup decision

Click the folder starter card, or just ask "work in my Projects/interviews folder". Nothing outside the folders you grant is readable — so grant the folder where the real files live. Click a folder's name under Access to open it in Finder/Explorer.

3 · The first real task — and what to expect while it runs

Read the three PDFs in this folder and write a one-page brief as brief.docx — keep the numbers in a table.

Turn results.xlsx into a 10-slide deck that argues we should fix mobile first. Speaker notes for my co-presenter.

While it runs: anything consequential asks first (approval cards); you can steer without stopping — type "use the December wave, not November" and it lands at the next safe step; drop files straight into the chat — from a granted folder they become @mentions worked on in place, from anywhere else they're copied visibly into the session's folder. Finished files land in your folder; the Files page keeps every deliverable in one place.

4 · Three keys — /, @, ⇧⇥

Key	What it does
/	App commands (/plan, /compact, /init, /model…), your saved commands, your skills
@	Point at a file in a granted folder — no path typing
⇧⇥	Cycle Plan → Ask for approval → Full access

For anything with stakes, hit ⇧⇥ into Plan first: Mimi proposes the whole approach, you approve or redirect, then it runs. Same gestures as Claude Code / Cowork / Codex — Settings ▸ Transfer guide has the full map.

5 · Teach it your way — once — instructions, skills, memory

• Instructions (Settings ▸ Instructions, or AGENTS.md/CLAUDE.md in the folder): standing rules. "Reports in UK English. Stats always with effect sizes."
• Skills (Settings ▸ Skills): package your brand rules or methods once; every deck and doc comes out in them without being asked. Browse 8,400 community skills, read a skill's real instructions before installing, import from ~/.claude/skills.
• Memory (Settings ▸ Memory): what Mimi noticed and kept — review, edit, delete.

6 · Connect where you work — Slack, Qualtrics, QualiTaTi data, 25+ more

Settings ▸ Connectors: Slack (tag Mimi in a channel, work comes back in-thread), Gmail/Outlook, Calendar, Drive, GitHub, Jira, Notion, Canva — and Qualtrics: read a questionnaire so Q4_1 becomes a real question, pull responses as CSV or labelled SPSS .sav, with your approval per download. Your QualiTaTi research data works the same way. Anything else speaks MCP.

7 · Make it run while you don't — automations and the Inbox

Every Monday at 8, read the new files in field-notes/, and put a one-page weekly summary in reports/.

That becomes an Automation — local, full transcript, and anything needing a decision waits in your Inbox instead of guessing.

8 · A good week, in five asks

1. "Work in this folder. Profile every .sav in it and give me a data dictionary as a Word doc."
2. "Package our brand guidelines into a skill."
3. "Turn the wave-2 findings into a 12-slide deck for the steering committee." (in Plan mode)
4. "Pull the December survey from Qualtrics as SPSS and test whether satisfaction differs by channel — effect sizes, not just p-values."
5. "Every Friday at 4, summarise this Slack channel's week into a memo in reports/."

By Friday: a data dictionary, an on-brand deck, a real analysis, and a standing automation — every file on your disk, made with your keys, under your approval.

Privacy

MimiWork is fully local: the agent loop, your conversations, connector tokens and model keys all live on your machine. There is no vendor cloud in the loop — no hosted sign-in, no OAuth broker, no relay. Connectors use your own credentials (or a vendor's own local OAuth); your data leaves the machine only through the model and integrations you choose.

Quality

The test suite is part of the product: 1,507 backend tests cover the agent engine, tools and connectors — including that every Office tool registers a schema a real provider accepts, that code execution is approval-gated like the shell, and that a scanned PDF is flagged rather than silently summarised from nothing. The desktop app carries 230 GUI unit tests and 150 hermetic Playwright e2e tests (mocked backend, no network), with a separate live suite for the real thing. ruff and eslint gate CI alongside them.

Build from source

┌────────────────────────────────────────────────┐
│               MimiWork desktop app             │  native shell + GUI (Tauri + React)
├────────────────────────────────────────────────┤
│           local agent server (Python)          │  engine · tools · connectors
├───────────────┬────────────────┬───────────────┤
│  your files   │   your tools   │  your model   │  everything runs with your keys,
│  & terminal   │ 25+ connectors │  any provider │  on your machine
└───────────────┴────────────────┴───────────────┘

Prerequisites: Python 3.10+, Node 20+, Rust via rustup, and cmake (brew install cmake).

git clone https://github.com/lanceyuu/mimiwork
cd mimiwork

# 1. One-time bootstrap — creates the Python venv at .venv
bash packaging/setup_dev_env.sh

# 2. Start the local agent server
.venv/bin/openworker-server --cwd ~/some/project --port 8765

# 3. In a second terminal, start the UI
cd surfaces/gui
npm install
npm run dev        # browser UI on the Vite dev port

For the full desktop app, replace step 3 with npm run tauri dev; build an installable DMG with packaging/build_dmg.sh.

Directory	What's in it
coworker/	Python backend — agent engine, model providers, office/PDF/image/analysis tools, connectors, MCP client, automations
surfaces/gui/	Desktop app — React UI + Tauri shell
stt/	Speech-to-text sidecar (Rust) for voice input
packaging/	Installer builds (macOS DMG, Windows), dev bootstrap
docs/	Design specs and decision logs
tests/	Backend test suite

Acknowledgements

MimiWork is a fork of OpenWorker by Andrew Ng and contributors, built on aisuite. The repetition-guard detection (stopping a looping model before it burns your credits) and the steer-while-running interaction are adapted from ideas in FrontierAgent by Apodex AI (Apache-2.0); the search deduplication and dead-end guidance in the web tools adapt the rollback observers of AgentHarness (Apache-2.0) to a desktop where the user pays per model call. The markdown /command files with $ARGUMENTS and the plugin-style tool hooks follow opencode's conventions, and the /, @ and Shift+Tab gestures deliberately match Claude Code and Claude Cowork so the muscle memory transfers. This fork repositions the tool for knowledge work — Office deliverables, PDF and image handling, and a statistics-aware data-analysis toolchain — and carries its own desktop refinements. Internal binary and package names (coworker, openworker-server) keep their upstream names for compatibility.

The full list of projects, skills and libraries MimiWork learns from or ships with — and what was taken from each — is in CREDITS.md, and in the app under Settings → About.

License

MIT — see LICENSE. Original work © 2024 Andrew Ng; modifications © 2026 MimiWork contributors.