The AI we build, and the AI we run on.

Zehntech builds AI agents, MCP servers and AI products, and uses them to run its own company. Scroll to see the work, in numbers.

    Scroll
    0
    AI agents
    0
    Skills written
    0
    MCP tools
    0
    MCP servers
    0
    Products
    27 AI projects2 in-house models8 model providers3 products built by AI agents
    Agents

    50 agents, each with one job.

    Every agent has a defined role, a scope and a budget. None of them writes to your data without a person confirming it.

    Scopedefined tools only
    Budgetcapped per run
    Writesperson approves
    How we build

    Three rules for every Zehntech AI system.

    01

    Agents with limits

    Every agent has a scope, a budget and an approval step, and a person signs off on anything it writes.

    02

    MCP first

    Models connect to real systems through MCP servers. We load only the tools a customer needs.

    03

    Spec before code

    Every AI project starts from a written spec and passes gates before release. We wrote our own Claude Code skill for it.

    Inside Zehntech

    We run Zehntech on our own AI.

    These AI systems are in daily use at Zehntech. As a result, the same patterns go into the work we build for customers.

    Market research

    HyperRadar watches eight sources. Perplexity finds what is happening, Claude works out what it means, GPT-4o drafts the output, and a person approves it.

    23 skills
    Sales

    Agents qualify leads and draft proposals, cover letters and replies. However, the sales team reviews every one before it is sent.

    13 agents
    Marketing

    Agents produce app-store GIFs, demo videos, listings and social packs for our Odoo apps.

    7 agents
    Quality

    Every Odoo app passes a 44-check scan and an AI code review before it is published.

    44 checks
    Engineering

    349 of our repositories are set up for Claude Code. Dictio and Markview were built entirely by AI agents.

    349 repos
    Research

    Our own models, and the ones we work with.

    7 ants

    ANT language model

    Seven nano models of 10K to 500K parameters, trained from scratch. Each does one small task well, on a CPU.

    1.5B

    Odient LM

    A small language model tuned for Odoo tool-calling. Runs on-premise, so data stays with the customer.

    5 engines

    AI Manufacturing Intelligence

    Predictive maintenance, defect prediction, OEE, demand forecasting and procurement inside Odoo 19. It runs without external APIs or a GPU.

    1 worktree per task

    FlowForge AI

    Agent orchestration where a master agent proposes and a deterministic engine enforces. Each task runs in its own git worktree.

    OpenAIAnthropicGoogle GeminiAWS BedrockOpenRouterPerplexityOllamaWhisperKit