About

An AI workforce,
with a human in the loop.

We build custom AI workers, operations automation, and AI integrations — and we run our own studio on exactly the same systems. Not an agency that added an AI practice. We built the systems first.

You get the operator
and his machine.

Crowork exists because most "AI for business" is a thin wrapper with someone else's assumptions baked in, sold by people who do not run it themselves. We took the opposite path. We build AI workers and automations for clients — and we operate our own studio on the same kind of systems, end to end.

When we say a supervised AI worker can hold a real role, it is because we have one holding ours. Every build is owned by someone who has run these systems in production. We supervise by design, because we have seen what unsupervised agents do.

We build in the open where it counts. The self-healing loop that lets our workers learn from their own failures (agent-self-improvement) and the gate that stops an AI from treating a question as a fact (speech-act-memory-gate) are both public and inspectable. We would rather earn trust by showing the work.

Talk to the builder, not a sales layer
Every tool battle-tested on our own ops before it ships to you
Flat, predictable pricing — never per token, per call, or per message
Open-source reliability tooling — we publish what we build

Direct contact

hello@crowork.ai

We respond to every message. Usually same day.

crowork_model.ts
// The Crowork model
const crowork = {
  workforce: "AI agents",    // the machine
  oversight: "human operator",// the ghost
  client_contact: "direct",
  account_managers: false,
  junior_handoffs:  false,
};
// What you get:
"the operator + his machine"

Human operator in the loop — every build, every output, every deployment.

Audit first, sell second

Every engagement starts with a free audit. If there is no clear improvement, we say so.

Honesty over hype

Confidence comes from demonstrated outcomes, not invented metrics.

Self-improving systems

The system gets better every day.

Most automation degrades over time as processes change. Ours does the opposite. Every worker we deploy learns from its own bugs, writes new skills for repeating tasks, and proposes fixes under human supervision — with the full loop running in production, not in a demo. The code that makes this work is published open-source so you can inspect it.

Failure loop

Learns from its own bugs

When a worker hits an unrecognised failure, it logs the error pattern, drafts a fix, and queues it for review. Approved fixes are merged back into the worker's live skill set — so the same failure doesn't happen twice.

Skill layer

Writes and ships skills

Tasks that recur three or more times get extracted into reusable named skills. The self-improvement repo (agent-self-improvement) is the working loop we publish publicly.

Memory gate

Keeps memory honest

A speech-act gate at memory ingest blocks questions from being stored as facts and guesses from being treated as verified data. The full implementation is open-source: speech-act-memory-gate.

Human-gated

Supervised by design

Self-improvement doesn't mean autonomous. Every proposed code merge, skill addition, or config change passes a human gate before it reaches production. The worker proposes; the operator decides.

agent-self-improvement speech-act-memory-gate

Both repos are public and inspectable. We build in the open where it counts.

Infrastructure & security

Built to be trusted.

Every worker we deploy runs on Fly.io — distributed, always-on infrastructure with EU-region deployments as the default. Persistent state lives in Supabase with encryption at rest and row-level security. Neither platform requires us to store your data outside the EU.

API keys are stored in environment secrets, never in plaintext in config files or databases. Unlike other AI agents, we don't store your keys in plaintext. That's not a marketing line — it's a design constraint we enforce at build time.

GDPR compliance is structural, not a checkbox. Data minimization is built into the memory gate. You can request a full data export or deletion at any time — and the system can execute it, not just acknowledge the request.

GDPR compliant Encrypted at rest EU data residency

Powered by

Anthropic Claude Supabase Google Workspace Telegram · MTProto n8n WhatsApp Fly.io OpenAI · Codex

Infrastructure: Fly.io (EU region, always-on). Persistent state: Supabase (encrypted at rest, row-level security). Messaging: Telegram MTProto + WhatsApp (Baileys). Workflow orchestration: n8n. All API keys held in runtime secrets, never in application config.

How we work

Operating principles

Applied to every engagement — including our own operations.

Build

Fast, no handoffs

Every build is owned by someone who has run the system in production. No ticket-routing, no junior hand-offs. Speed comes from knowing the failure modes before writing the first line.

Safety

Supervised by design

Human oversight is built in from day one. Every action is logged; edge cases go to a person, not a guess. An unsupervised agent acting on your operation is a liability — we treat it that way.

Pricing

Flat and predictable

Fixed scope, fixed price, flat monthly to run. Never per token, per call, or per message. You can forecast the cost like any other fixed operating line.

Core

AI-native, not AI-adjacent

We run our own operation on the systems we build, and open-source the reliability tooling beneath them. These are not demos — they are production systems.

Trust

Honesty over hype

If AI is not the right fit, we say so in the free audit. We would rather lose a project than sell you filler. Confidence should come from real, demonstrated outcomes.

QA

We test the real thing

We reproduce the failure first, then fix it, then verify the user-visible result actually happened. "It compiles" is not "it works." Every release passes human review before it goes live.

What's next

See the work.
Understand the process.

The best way to understand how we operate is to read what we have built. Every case study is a real engagement — anonymized, but specific.

If you want to understand the engagement model, the process page walks through discovery, build, and run in plain language.