For regulated mid-sized companies · Stuttgart, Germany
We build software for your business processes. It records who decided what, and on what basis. Any later change to that record can be detected.
New projects start from November 2026. We are happy to talk now and show you Approval in a short call.
An example: Approval, one of our applications
A department requests software. Approval reads five vendor documents, drafts both assessments and puts them in front of your officers.
This shows the process used in Approval, with fictional requests, documents and figures.
The same division of work fits your own processes, as a new application or inside software you already run.
Custom software
Security, data protection and the review of AI proposals are already built into our template. Coding agents add your process on top of it; automated checks send back whatever breaks its rules, and we review what passes. It runs on your infrastructure, you own the code, and there is no platform subscription.
AI oversight
For software you already run: a model prepares a proposal, a person reviews and signs it, and a hash-chained ledger records the decision. Open-source libraries in your own stack.
Custom software
Writing code has become cheap. Coding agents are fast, but left to themselves they pick the wrong tools, skip requirements and leave gaps that show only in production. So no agent of ours starts from an empty folder. Every agent uses our template, libraries, instructions and automated checks. We maintain all four.
@octabits-io/createEvery application starts with row-level security on every table, sign-in through your identity provider, AI steps a person approves and a log from which each change can be undone. The agent adds your process to this template.
@octabits-io/*Queues, encryption, data-subject requests, events and the audit log come from our open-source packages, each with its own tests. The agent wires them in instead of writing its own. All of them run on your Postgres, so none adds a sub-processor.
AGENTS.md · skills/The template gives the agent instructions: how to add a table, how to change one that already holds data, how to add an AI step. Each recipe ends with a command to check the result. Claude Code and Codex read the same files.
pnpm guardWhen the agent ends its turn, the guard checks the rules: every table scoped, organisations isolated from each other, migrations safe on tables that hold data, no model call that bypasses the disclosure rules. What fails goes back to the agent. The same checks run before every merge.
We read every change that passes the checks before merging it. The agent writes most of the code; what goes into your application is decided by people.
Recorded development
We asked Claude Code, Anthropic’s coding agent, to build a customer management application from a fresh copy of our template. It needed companies, contacts and a company picker. Deleting a company had to preserve its contacts. It chose the data model and worked without intervention for just under nine minutes. The thirty-five-second excerpt is followed by a recording of the finished application.
Claude Code builds customer management on the octabits starter: the task, the skills it reads, the migration, the guard with 9 of 9 checks, its review, a smoke test that fails and one that passes, and its closing message.
The guard has grown since the Codex run, from 6 checks to 9: it now also replays every migration over tables that already hold rows. The run took 40 turns and cost $2.38, as shown at the end of the recording.
We gave a coding agent, OpenAI’s Codex, a fresh starter and one sentence: invoices, a list, and a button that summarises an invoice with the model we run in-house. It worked unattended for just under nine minutes. The thirty-second excerpt is followed by a recording of the finished application.
Codex builds invoices on the octabits starter: the task, the skills it picks, its question for the in-house model’s address, the guard with 6 of 6 checks, its own tests and its closing message.
No model was connected for this recording. The application used scripted demo responses, as noted by the agent and on the page. Three settings are needed to connect your own model.
How it works
Tell us how the work gets done, what data and documents you use, and who approves the result. A spreadsheet you use today is a useful starting point. In the first call, we outline how to build your application with our existing libraries.
Coding agents write most of the code inside our template. Automated checks send back whatever breaks its rules, and we review every change before it is accepted. You can use the first version within a week.
The application uses your servers, Postgres database and sign-in provider. Your staff review AI proposals, and the log records their approvals. You own the code.
AI oversight
Approval uses the oversight kit to review software requests, as shown in the diagram above. The model drafts assessments with source references. The officers review and sign them, and each decision is logged. We can integrate the same libraries into your existing software.
Available from November 2026
AI prepares the assessment for your officers to review and sign.
Approval helps regulated organisations assess and approve software. It collects vendor documents, extracts facts with source references and drafts security and data protection assessments. The responsible officers review, edit and sign the documents. Approval runs in your own datacentre on your Postgres database.
Pricing
You own the application code. There are no licence fees or limits on users and workflows. The platform packages will be released under the MIT licence.
Development
Fixed price
per application, in the written proposal
Maintenance
Monthly
a flat fee, no charge per user
For business customers. After a first call you receive a written proposal with scope, duration and price, usually within two working days.
Consulting
For teams with a system of their own: four engagements with a defined scope, duration and deliverable. The first call is free.
Open source
Approval and Reynt use our open-source libraries. We group them into two kits. The build kit handles row-level security, field encryption, data-subject requests, queues, events and telemetry. The oversight kit adds review of AI proposals and records the decisions. The packages run on your existing Postgres database. Each comes with tests and documentation explaining its uses and limits, and all of them will be released as open source under the MIT licence in November 2026.
octaflowDurable workflows on the Postgres you already have.Documentation →
Also maintained:cert-manager-webhook-porkbunWildcard certificates for Porkbun domains on Kubernetes.
Capabilities and automated checks
How we build
We explain the patterns we use, with examples from our code and the trade-offs involved. They apply across programming languages.
In production
We develop and operate Reynt, a booking platform with thirteen packages, four APIs and three frontends. AI coding agents write most of its code. We began publishing individual libraries after using them in production in Reynt.
Automated checks enforce the architecture boundaries, the database enforces key constraints, and integration tests run against a real database in CI. These checks support our review of agent-written code.
Reynt in numbers
Tell us briefly what you are working on and where you need help. Email is the fastest way to reach us; the person who would work with you will reply.