Runs on your machine · local or cloud models, your choice
The private AI assistant for every desk
It reads your documents and databases, does the work on your machine and hands back the finished file.
- 10-day free trial, no card
- Ollama or your own API key
- Windows · macOS and Linux coming soon
- Open bank_export_sept.csv
- Match 214 payments to invoices
- 3 mismatches flagged for you
- ✓ Month closed
Finance closes the month.
One app, every department
Twenty-six job roles ready to use — office and engineering desks alike — each a saved agent with its own instructions, its own tools and its own limits. Set one up once and the whole team can address it.
Work that comes back finished
Reads PDF, Word and Excel, answers from your own databases, researches the web, and produces a real .docx, .xlsx, .pdf or chart — generated locally, not by a cloud converter.
Your model, your machine
Runs local Ollama models or your own API keys. In local mode your AI requests stay on your network — enforced by a one-click lock.
Made for confidential work
Your account confirms who you are. Use a local model and your documents and prompts can stay on your machine.
Roles
One assistant, every desk
Each role is a saved agent — its own instructions, its own tools, its own limits. Pick the ones your company needs; a seat includes them all.
Front office
Drafts the correspondence, files the intake, and keeps the follow-ups from going quiet.
Finance & bookkeeping
Reads the invoices out of Excel and PDF, reconciles them, and writes the summary back as a real spreadsheet.
HR & people
Turns a policy change into the letters, contracts and onboarding packs that have to go with it.
Sales
Builds the quote, drafts the proposal, and answers “what did this customer buy last year” from your own database.
Customer support
Triages the ticket, drafts the reply, and writes the article so the next person does not have to ask.
Operations
Pulls the daily numbers from the system that holds them and puts the same report on your desk every morning.
Marketing
Researches the market, writes the copy, and produces the one-pager as a finished PDF.
Legal & compliance
Reads the long contract, marks what changed against the last version, and drafts the clause.
Data & reporting
Answers a question in plain language by querying the database, then charts it.
Project management
Keeps the plan, chases what is late, and writes the status update nobody has time to write.
Procurement
Compares the suppliers, checks the terms, and produces the order document.
Executive assistant
Reads the eighty-page pack and gives you the two paragraphs you need before the meeting.
Inside
What's inside
Five toolboxes in one app, all included in every seat — take the ones your job needs.
The everyday assistant
The work that fills a week in any department, done properly and handed back as a file.
- ▪Reads PDF, Word and Excel — the eighty-page pack included
- ▪Writes real .docx, .xlsx and .pdf files locally, ready to send
- ▪Draws charts from your own data
- ▪Researches on the web and cites what it read
- ▪Keeps a visible plan: what it intends to do, and what it has done
- ▪Stops and asks instead of guessing, whenever you want it to
- ▪Hands work to a colleague or another agent, and collects the report
The coding agent
For the technical desks: reads your repo, edits files, runs commands — and, with auto-approve off, stops at least before every step the app rates as dangerous, such as a force-push or dropping a table.
- ▪Multi-step agent loop with real tool use
- ▪File edits with a per-change diff and one-click rewind
- ▪Shell + terminal sessions, several at once, with live input
- ▪Risk-classified approvals — with auto-approve off, safe reads run and destructive commands ask
- ▪Read-only agents, and per-agent tool allowlists
- ▪Multi-agent teams: delegate, or orchestrate a whole task across them
- ▪MCP servers, so it can drive your own tools
- ▪Scheduled tasks that run unattended
The database IDE
A real client, not a chat box that guesses at your schema — and plain English for everyone who does not write SQL.
- ▪SQL console with the schema in context
- ▪Plain English → SQL, with the plan explained back
- ▪EXPLAIN warnings before a slow query hits production
- ▪Ask for an index and it reads the query to suggest one
- ▪ER diagrams generated from a live database
- ▪PostgreSQL, MySQL, SQLite and MongoDB
- ▪SSH-tunnelled connections to a remote host
- ▪Every statement risk-classified before it runs
- ▪Remote-server tools: monitor load, disk and GPU
Train your own model
Turn the work you already did into a model that fits how your company works.
- ▪A second model scores every reply, building a dataset as you go
- ▪Your corrections become preference pairs
- ▪Curate, de-duplicate and export to JSONL
- ▪LoRA / QLoRA fine-tune on your GPU — or a remote one over SSH
- ▪Generated Python pipeline: analyse → train → merge → evaluate
- ▪Test cases to verify a fresh build, then A/B it against the base
- ▪Import the result into Ollama and switch to it in the model picker
The everyday bits
The parts you notice only when they are missing.
- ▪Preview PDFs, images, audio and video without leaving the app
- ▪Code editor with formatting, linting and AI rewrite in place
- ▪Git panel: stage, diff, commit, branch
- ▪New-project scaffolding that knows your OS and installs the right way
- ▪Token and spend tracking per project and per session
- ▪Route chat, images, audio and video each to a different model
- ▪Team chat, in the app and on the web, with your bots in it
Autonomy you can bound
An assistant that can act, and a brake you control
Every action is classified before it runs. Auto-approve is off until you turn it on — on a new installation and for every new account — and while it is on, everything runs without asking, whatever approval rule a role has — except installing or removing a program on its card, which always asks. With auto-approve off, reading a file or a spreadsheet goes through; deleting, force-pushing or dropping a database table stops and tells you why it was flagged — and you decide per role how much rope it gets, from confirm-everything to fully autonomous. Either way, read-only agents can't change anything at all.
The data boundary is a visible switch, not a setting buried in a menu. If your account is not part of a team (and the machine is not a team bot), flip Local-Only and the app's own internet connections, such as AI providers, web search and websites in the browser panel, fail loudly, naming the host they tried to reach. Still reaching the internet: the daily update check with medacrew.com, and the programs the app starts for you or an agent — git, shell commands and package installs (except program installs through the app's install card, which Local-Only blocks), SSH, database clients and local MCP servers. Local-Only also stops the access check with medacrew.com, so no later than one day after the last check the app stops until you lift Local-Only and sign in again.
- Approval gates — With auto-approve off (the default), file writes, SQL, SSH and shell commands the app does not rate as read-only ask before acting — except for agents with their own, looser approval rule. The rating is automatic and can be wrong.
- Risk reasons — A prompt says what is dangerous, not just that something is.
- Local-Only lock — One click on an account without a team; blocks the app's own internet requests at the source. The daily update check and programs the app starts (git, shell, SSH, database clients) still go out. The access check pauses too: no later than one day after the last check the app stops until you lift the lock.
- Full audit trail — Every run, tool call and file change stored locally and exportable.
- Rewind — Reset, turn by turn, the text files the app changed with its file tools; what it newly created, documents and images included, is removed. Deleted and moved files, commands, database statements and sent messages are not brought back.
- Central revocation — Sign a device out from the web, for teams.
Models
Bring your own model
The app connects to the provider from your machine. Two of these need no key and no connection at all.
Ollama
Runs on your machine
Local, free, private. The default.
Claude
Your own API key
Anthropic Claude via API key.
OpenAI
Your own API key
GPT-4o, o-series. Needs an API key.
Google Gemini
Your own API key
Gemini models on your Google API key.
OpenRouter
Your own API key
One key, hundreds of models.
Groq
Your own API key
Very fast hosted inference.
LM Studio
Runs on your machine
Local OpenAI-compatible server.
Any OpenAI-compatible
Your own API key
Any OpenAI-compatible base URL.
Setup
Up and running in three steps
Download & install
Grab the app for Windows. macOS and Linux are coming soon.
Start your trial
Sign up in the app on first launch. Every feature is unlocked for 10 days — no card, nothing to choose.
Pick your seats
Before the trial ends, buy a seat per person and a bot seat per unattended runner. Manage both from the web.
Pricing
Simple pricing
10 days of the complete app, free and without a card. After that a seat per person — whatever their job — and a separate seat for each machine you leave running on its own.
Free trial 10 days
€0 for 10 days
The whole app, every feature, for 10 days. No card, so nothing is charged when it ends — it simply stops until you pick a seat.
- Every feature unlocked, nothing held back
- No card, nothing to cancel
- Bring your own model or run local
- Keep whatever you set up when you subscribe
User seat
€9 / person / month
Final price · no VAT (§ 19 UStG)
Billed monthly, cancel anytime
- One per person, in any department
- Every feature of the app, nothing locked
- All twenty-six job roles, and any you write yourself
- Unlimited projects & sessions
- Shared team chat, in the app and on the web
- Assign work to a colleague’s agent (managers)
- Central control — revoke a device or seat any time
Bot seat Unattended
€49 / bot / month
Final price · no VAT (§ 19 UStG)
Only if you run machines unattended
- One per unattended runner
- Runs work with nobody watching — reports, CI, deploys
- Answers as @handle in team chat
- Replaces that account’s user seat, not added to it
- Revoke it any time and the seat frees up
Final prices. As a small business under § 19 of the German VAT Act (UStG), we charge no VAT. Secure payments by Stripe.
Questions
Questions
Reviewing this for a company? The trust & compliance page covers GDPR and the EU AI Act.
Keep the work in the building. Keep the AI anyway.
Install in a minute and try everything for 10 days — no card. Add seats as more desks want one.
Download MedaCrew