Decision Sovereignty

Memory + Decision Engine
for AI in production.

Typed memory, audited decisions, evolving with use.

Start building Book a decision review

zunft@localhost:8000

live

Decisions generated1,284

HITL approvals412

Human overrides38

Avg agency-health0.87

Active kernels26

▸Deal #4421 → 3 options surfaced → approved.

▸Lead #7182 → qualifying decision → routed to reviewer.

▸Ontology "sales_v3" → 2 new slots authored → validated.

Enterprise-Grade Security & Compliance

GDPR

EU Compliant

Sovereign Cloud

Deployable

SOC 2

In progress

ISO 27001

In progress

How it works

How you build with Zunft.

Three steps. Each verb is a typed surface. Each step ends with a persona that maps to the buying motion — bottom-up adoption to top-down approval.

5 min

author

Solo Dev

Prompt Claude Code to author a typed decision kernel via MCP. The kernel authoring engine emits a typed class and an audit row.

1 day

decide

Tech-Lead

Wire the kernel into your agent loop. Every invocation produces a typed ChoiceArchitecture — options, trade-offs, provenance.

1 week

audit

CTO / Compliance

Every decision is a row in the audit trail. HITL gates route human approvals. Overrides feed the calibration loop.

A decision is the unit of work — not a commit.

Authored with the IDEs your engineers already use. Runs on the protocols your agents already speak.

Authored with

Claude CodeCursorClineCodex

Others gated on MCP stream-compat verification.

Runs on

HTTPMCPPython SDK·LangGraphPydantic AIGoogle ADK·Claude SDKOpenAI SDK·Windmill

How it ships

From CRM optimization to decision engineering.

The CRM optimization trap

Sales orgs keep adding fields, workflows, and integrations to systems built for record-keeping, not decision-making.

The decision-engine shift

What if the system optimized decisions, not records? Typed sales ontologies, authored decision kernels, audited HITL approvals.

Decision-Orchestration

Sales Ontologies

Compute

What ships today

Currently shipping with design partners across DACH industrial-services commercial use-cases. Production deployments under NDA.

Apply to be a design partner

Built against real industrial commercial use-cases. Compute + Sales Ontologies + Decision-Orchestration. Open the architecture →

Quickstart

Author your first kernel in 5 minutes.

// ~/.config/claude-code/mcp.json

{
  "mcpServers": {
    "zunft": {
      "transport": "streamable-http",
      "url": "https://api.zunft.ai/mcp/tenant/stream"
    }
  }
}

First call opens a browser to sign you in to your tenant. Token is cached locally.

Then in Claude Code:

“Author a lead-qualification kernel for our industrial sales pipeline. Include a HITL gate and emit the audit row.”

zunft@api.zunft.ai/mcp/tenant/streamtenant: ind-sales-pilot

→ zunft.author_kernel
    prompt   "lead-qualification · industrial sales pipeline"
    gate     hitl
    emit     audit

✓ kernel authored
    class    lead:QualifyingDecision
    slots    lead_score · fit_signals · risk_flags · recommendation
    HITL     gate active  ·  1 reviewer required
    audit    lead-qual-2026-06-25-001  ·  sha256 7a3f…

→ ready  ·  next: zunft.generate_decision

2026-06-25 11:42:18 UTC423 ms

5 min to first authored kernel.

Works with any MCP-capable client (Cursor / Cline / Codex — verified compat varies).

How we price

Mechanism visible. Amount is conversation.

The meter

We price by typed decisions generated. One ChoiceArchitecture emitted = one billable decision.

What's included

Memory storage · audit trail · HITL gates · tenant isolation · OAuth — included in the meter.

The conversation

For volume estimates, sovereign cloud (EU/DACH), on-prem deployment, multi-year commit — talk to us.

Apply to be a design partner

About

Decision sovereignty for AI in production.
Built so your AI can be reviewed.

We engineer typed runtime for multi-agent systems — ontology authoring, decision generation, audit, HITL. Tenants own their kernels. Operators don't see your data. Reviewers see every choice.

Built and operated by a small senior team in Munich.

Supported By

Google Cloud

Startup Program

Xpreneurs Graduate

Germany's top incubator

Built by serial founders.

Ready to make your AI CTO-fest?

Book a decision review — we look at your agent's outputs and show you how Runtime reviews them.

Book a decision review Read how it ships

Memory + Decision Enginefor AI in production.