What Happened
Soapbox, a decentralized social media software company, has open-sourced an AI accounting agent called Sheila that handles the full lifecycle of contractor payments. The project was shared on Hacker News on March 6, 2026, and the source code is available on GitLab under the AGPL license.
Sheila is not a SaaS product with a polished UI. It is a collection of over 50 granular TypeScript scripts, each performing a single task: checking bank balances, sending ACH or wire payments through Mercury, processing crypto payments via Kraken and Lightning Network, generating invoice PDFs, uploading to Google Drive, and submitting expenses to OpenCollective. It also handles P&L tracking and 1099 report generation.
What makes the architecture interesting is how it chains these scripts together. Instead of hard-coded multi-step workflows, Sheila reads a 600-line AGENTS.md file written in plain English that describes how to combine the individual tools. The agent runs interactively in a terminal, and the operator watches operations before transactions execute - there is a human in the loop before money moves.
The system was built using OpenCode, an AI coding platform, and integrates with Google Sheets as its ledger rather than a traditional database.
Why It Matters
Most AI accounting tools - QuickBooks, FreshBooks, Xero - are adding AI features as incremental upgrades to existing platforms. Sheila takes a fundamentally different approach: it is an AI agent that was designed from the ground up to run accounting workflows autonomously, with human approval at the payment step.
For small companies and freelancers managing a handful of contractors, the current process is painful. You receive invoices by email, manually enter them into accounting software, initiate payments through your bank, generate tax documents at year-end, and file everything in shared drives. Sheila collapses all of that into a single agent workflow.
The crypto payment support is notable too. For companies that pay international contractors in cryptocurrency (increasingly common in tech), having ACH, wire, and crypto payments handled by the same system removes a layer of manual coordination.
Our Take
Sheila is more proof-of-concept than production-ready tool for most teams, but the architecture is worth studying. The idea of describing workflows in a plain-English markdown file rather than coding complex automation pipelines is exactly the pattern we expect to see more of as AI agents mature.
The 50-script toolkit approach is smart. Each script does one thing and can be tested independently. The AI layer handles orchestration, not execution. That means when a payment API changes or a new service needs to be added, you write one new script rather than refactoring an entire application.
The AGPL license and terminal-based interface will limit adoption to technical users, which is fine - this is clearly built by developers for developers. But the pattern it establishes - small composable tools plus an AI orchestration layer reading natural language instructions - is one that commercial accounting platforms should be paying attention to. The gap between what Sheila does with 50 scripts and a markdown file and what enterprise accounting software does with millions of dollars in development is uncomfortably small.