Our team was spending $700/month on AI. Claude subscriptions, GPT-4 API calls, experiments that ran up bills before we noticed.
Now we spend $10-20/month for the whole team.
We didn't stop using AI. We stopped overpaying for it. And we built a desktop app that lets anyone — not just developers — run AI agents on their own computer, with their own files, using whatever model they want. Local, open-source, cheap API, or premium — pick when the task demands it.
This is Zosma Cowork.
The Problem With AI Right Now
If you're non-technical and you want AI to actually do work — process invoices, generate reports, analyze spreadsheets — you have bad options:
- Web chatbots (ChatGPT, Claude) — They can't touch your files. You copy-paste, they guess. $20/month each, and you still do the actual work.
- Coding agents (Claude Code, Codex) — Powerful but terminal-only. You need to know command line, install things, configure API keys. Not for most people.
- SaaS AI tools — $50-200/month per tool, each for one specific task. They add up fast.
None of these let you use your own local models. None of them let you pick cheap providers for easy tasks and save the expensive models for hard problems. And none of them are open source.
What We Built Instead
Zosma Cowork is a desktop app. You install it, add your API key (or point it at a local model via Ollama), and start giving it tasks. No terminal. No TUI. No subscription.
It runs on anything. We support OpenAI, Anthropic, Google, Groq, Together AI, and local models via Ollama or LM Studio. For sensitive work (financial data, customer records), we run Qwen 2.5-32B locally — the data never leaves the machine. For everyday tasks, we use cheap API providers that cost pennies.
Our actual costs: $10-20/month for the entire team. We use local models for finance, cheap API for design and content, and only use Claude for the hardest problems (about 5% of usage).
Three Things It Can Do (Recorded in One Take Each)
Every demo below was a single shot. We asked the agent and started recording. No pre-trials, no editing.
1. Design a Mobile App
The ask: "Design a food delivery app."
The agent wrote HTML, CSS, and JavaScript for a multi-screen prototype — home screen, menu, cart, checkout. It iterated on layout issues it noticed itself. No human touched a design tool.
2. Process Invoices Into a Spreadsheet
The ask: "Extract data from these PDF invoices and compile it into Excel."
The agent read the PDFs, identified fields, and wrote a formatted .xlsx file. What takes 15-30 minutes manually happened in under a minute.
3. Generate a Financial Report From a Bank Statement
The ask: "Take this bank statement and generate a readable report."
The agent categorized every transaction, calculated totals, identified spending patterns, and produced a structured report with summaries.
Why This Works for Non-Technical Users
We built Zosma Cowork because the existing agent tools (Claude Code, Codex, pi) are all terminal-first. They assume you're comfortable with a command line, environment variables, and config files.
Most people aren't. And they shouldn't have to be.
Zosma Cowork wraps the same agent capabilities in a desktop GUI. You type what you want done. The agent figures out the rest — which files to read, which tools to use, which steps to follow. You can watch it work or let it run in the background.
The underlying agent system (pi coding agent) has over 700 extensions on npm. You want database access? Install an MCP server. Need Google Sheets integration? There's an extension. But you don't need to know any of that to use the app. The extensions just make it more powerful if you want to go deeper.
No Lock-In
Everything we built is open source under MIT. The desktop app, the agent sidecar, the extension system — all on GitHub.
- Use any model provider
- Run local models for sensitive data
- Switch providers per task
- Take your data with you
- Modify the code if you want
Getting Started
- Download the latest release
- Add your API key (or set up Ollama for local models)
- Start giving it tasks
No subscription. No terminal required. No lock-in.
Requirements: macOS or Linux (Windows coming).
Star the repo on GitHub · Watch all demos · Join the community
Top comments (1)
Routing the right task to the right model is where the real savings sit, and once you start mixing local OSS models for sensitive data, cheap APIs for the easy work and premium only for the genuinely hard problems, the cost story basically rewrites itself
It's the math every team running production AI is going to be forced to do the moment finance starts looking at the bill, and keeping the data on the machine makes the security and compliance conversation a lot shorter too.