T-Matrix
EN
Get a quote

Stellwerk · Early access

The control room for your AI agents.

Claude Code, Codex, Cursor and your own bots in one place, with shared memory, their own keys, spending limits, personal data masked before it reaches a model, and a journal for the AI Act.

Book a demoPricing

The problem

Agents everywhere. No control, no record, no budget.

  • Agents are everywhere

    Every team runs its own assistants and coding agents, each with its own keys, memory and habits.

  • No overview of spending

    Model costs arrive as a surprise at the end of the month, with no answer to who spent what.

  • Data and the AI Act

    Client data can end up in a model outside the EU, and nobody keeps the record the AI Act now expects.

Stellwerk is German for a railway signal box: the room where points are switched and signals set, so every train runs on its own track and none collide. That is what the product does for your agents.

Our product

Stellwerk. The control room for your AI agents.

US model: blocked for client dataEU-hosted modelAgentyour AI agentMaskhides personal dataGatewayroutes and limitsGuardhuman approvalJournalAI Act record
MASKWhat the AI model seesInvoice for Lena Huber, IBAN DE89 3704 0044 0532 0130 00
GATEWAYClient data only to EU-hosted modelsTeam budgetwarn 80% / stop 100%
GUARDDeploy to production? Waiting for a human Approved by a human
JOURNAL10:42:03 agent.request10:42:03 mask: 2 fields10:42:04 gateway: eu model10:42:09 guard: approved
  1. WorkspaceAn agent asks for help. It works with its own key, shared memory and MCP.
  2. MaskNames, e-mails and IBANs become placeholders before anything leaves your company.
  3. GatewayClient data goes to EU-hosted models only. Spending stops at 100% of the budget.
  4. GuardRisky actions wait for a human. Deploys, payments and outgoing mail need approval.
  5. ComplyEvery step lands in the journal and becomes your AI Act record.

Modules

Five modules on one track.

Connect an agent

Connected in about a minute.

Any agent that speaks MCP can join: Claude Code, Codex, Cursor and your own bots.

  1. 01

    Create the agent

    Register the agent in Workspace. It gets an owner and its own key.
  2. 02

    Add one address

    Point the agent's MCP settings to Stellwerk. One key replaces the keys of every model provider.
  3. 03

    Work as usual

    The agent sees its projects, tasks and shared memory. Every request goes through Mask, Gateway and Guard.
Example: Claude Code
claude mcp add --transport http stellwerk \
  https://stellwerk.t-matrix.net/api/mcp \
  --header "Authorization: Bearer <your-agent-key>"

Security

Built for companies that answer to the GDPR and the AI Act.

  • Hosted in the EU

    Stellwerk runs on servers in Germany. Enterprise customers can run it in their own cloud or on their own servers.

  • Companies strictly separated

    One system, many customers: no company can see another company's data.

  • Your own model contracts

    You connect your own contracts with OpenAI, Anthropic, Google or Mistral. We route and count, we do not resell tokens.

  • Watches agents, not people

    No emotion recognition, no productivity scores for employees. The journal records what agents do, with access by role.
Security for IT teams

Pricing

Priced per company, not per token.

  • Team

    €149/month

    Up to 5 agents

    • Workspace
    • Mask

  • Business

    €490/month

    Up to 25 agents

    • Everything in Team
    • Guard
    • Gateway
    • Single sign-on (SSO)
    • Audit journals

  • Enterprise

    from €1,900/month

    Your cloud or your servers

    • Everything in Business
    • Install in your cloud or on your servers
    • SLA
    • Comply included

Prices excl. VAT. Comply is included in Enterprise or available as an add-on for €190 a month. You pay model providers under your own contracts: we add no markup on tokens.

FAQ

Good to know.

Is Stellwerk available now?
Our own internal version already runs our agents every day. The product for customers is in early access: book a demo and we will set up a workspace with you.
Do you charge for model usage?
No. You pay model providers under your own contracts. Stellwerk routes requests and counts spending, without a markup on tokens.
Does Guard monitor our employees?
No. Guard records the actions of agents, not people. There is no emotion recognition and no scoring of employee performance, which matters in Germany, where tools that can assess employees need the works council's agreement.
Is Comply legal advice?
No. Comply is a tool with templates and records. In Germany only a lawyer may assess your specific legal case. We can arrange an individual review with a lawyer.
Which agents can connect?
Any agent that speaks MCP, for example Claude Code, Codex and Cursor, plus your own bots.

Tell us what
to build.

Get a quote
hello@t-matrix.net

We reply within one business day.