How Translucent Builds an Accounting Factory for 15,000 Businesses

Translucent began in 2022 as an accounting platform for businesses with multiple entities. More than 15,000 businesses have used it. The founder, Michael Wood, previously co-founded Receipt Bank (now Dext), which grew to serve more than 1M businesses. His thesis with Translucent was that every accounting tool is just a narrow slice of the core ledgers. He believed that building great databases first would let him build tons of apps on top.

That foundation worked well for agents. Translucent already had ClickHouse and Postgres databases that reconstructed ledgers from small primitives, calculated metrics quickly, and exposed the data through read and write APIs. An agent could work from a structured view of the books.

Using Mastra, Translucent built an accounting factory on top: a shared system of tools, skills, workspaces, and human checkpoints that can support many different accounting jobs.

Why Mastra

Translucent had already built an earlier AI assistant, using Langfuse, but decided to replace it with Mastra when agents became central to the product. The framework was suggested by Jürgen Brandstetter, who joined Translucent as CTO. Jürgen had followed Mastra since its early Hacker News launch.

Mastra's TypeScript foundation matched Translucent's stack, and the team appreciated its open-source nature. More importantly though, Mastra's primitives matched Translucent's product direction. They didn't want isolated agents, but rather one harness whose memory, tools, skills, and orchestration could improve together.

"Mastra was philosophically aligned with what we were trying to do. And it's open source, which means we don't have the same risk of getting locked in." — Deane Barton, CEO and Co-founder

Combining agent judgment with deterministic tools

The biggest opportunity for the agentic system was in the work between data collection and client advice: posting journals, maintaining schedules, reconciling balances, and preparing reports. Easy to understand processes, but the inputs vary a lot, so strict automation won't work. Plus, accountants are used to delegating work and reviewing it, so an agent helper, in a way, fits existing behaviours.

Jürgen describes the system they built as an accounting agent factory. They have one agent that can perform over 20 different accounting tasks thanks to tools and skills. The agent can use roughly 60–70 read and write tools. Many expose APIs that Translucent had built earlier; others give the agent the specialist context that the original application didn't need.

The tools keep complex data access deterministic. The agent only asks for the accounting view it needs. For larger jobs, it can query data with SQL and generate an Excel output. It doesn't have to load tens of thousands of rows into the prompt.

"The agent does the judgment, but it doesn't do the mathematics. The tooling is there, and suddenly we don't have problems with context because the return value can be one line." — Jürgen Brandstetter, CTO

Skills sit one level above the tools and give an agent the procedure.

Prepayments are a good example of how tools and skills work together. The accounting rule is deterministic, but the evidence isn't. One customer attaches an invoice showing an annual subscription, another only provides a bank label. The model interprets the available evidence, identifies what's missing, and follows up with the accountant when it cannot make a safe decision.

Turning chat threads into accounting jobs

The first version of the product still looked like a familiar AI chat. Customers understood the answers were grounded in their real accounting data, but they didn't see a major productivity advantage over general-purpose tools.

The problem was the interaction model. Accountants don't sit down to have a conversation. They open a list of jobs due this week, each with a deadline and a defined scope, and work through it. Open-ended chat is a poor fit for work that is already known and plannable. The agent needs to be paired with workflow management, that's built around how accounting work is actually structured: recurring jobs, per-client deadlines, review and sign-off.

The solution turned out to already exist inside Translucent's Mastra-based system. A thread could be a task. Translucent attached accounting-specific metadata — deadlines, states, ownership — directly onto each thread, and a chat became a job without any new infrastructure.

From there, the idea kept expanding. Some accounting work isn't just one task; it's actually eight tasks, performed every month. So a thread could become a checklist, with steps running sequentially or in parallel.

The next challenge was supporting larger orgs. Someone comes in with a 60-entity business and an Excel file, wanting to post seven different journals to each one. Each one needs a small bit of enrichment. Mastra's workspace and ephemeral sandbox capabilities give each job somewhere to hold files, run scripts, and produce artifacts.

As the product expanded from a single accountant to whole teams, Translucent adopted Mastra's multiplayer support. Now several people can work inside the same job and see the same human-in-the-loop request.

"Even though we're all in the AI bubble, our customers are not. They have work that needs to be done every week and they have limited time to play around. What's interesting is that this product intuitively makes sense to them. This concept of multiplayer harnesses, with core primitives around it, we don't explain any of that tech stuff to people. They see it and they get it." — Deane Barton, CEO and Co-founder

Scaling the system in production

Translucent's agents are already being used by accounting firms and businesses in private access, with new firms joining through its early-access waitlist each week.

Real usage exposed new challenges like making long-running work dependable for many people at once. The team had to solve issues around streaming, restoring human approvals, tracking background agents, synchronizing shared threads, and scaling beyond a single server. Mastra's early durable-agent implementation wasn't enough for Translucent's needs, so the team temporarily added its own durability layer. At the same time, Mastra keeps improving.

"Every three days, I have a ticket that looks at Mastra's changes and our changes and rips out everything we don't need anymore. It makes our platform more stable because it follows the way the Mastra team intends it to run, not the way we glued it together around a bug." — Jürgen Brandstetter, CTO

Translucent also keeps improving individual tools and the accuracy of specific jobs based on feedback from users. Next, they want to connect more signals from customers' systems, wake agents when something changes, and use isolated subagents to process multiple companies in parallel before returning a combined result for review.

The goal is not to replace the accountant's judgment but to give accounting teams a way to delegate the work that happens before that judgment, without losing visibility or control.