Kizma
Become a design partner

The control layer for AI software work

AI Software Teams.
Controlled.

Claude Code, Codex, Cursor, and other coding agents do the work. Kizma controls what they're allowed to do.

Become a design partnerSee how it works

Design-partner access is coordinated directly. Kizma does not execute agent actions.

  1. Agent request

    Claude Code

    command_executionreported by the runtime

  2. Kizma control boundary

    Read-only inspection

    git-statusclassified by Kizma, not by the runtime

  3. Control decision

    Human approval required

    No automatic authorization exists.

  4. State

    Awaiting human decision

    Execution not authorized.

One control loop

Authority before action.

Kizma sits between an agent's request and the next permitted step, keeping task scope, policy, human decisions, and evidence connected.

Agent capabilityAgent authority

Capability is not authority.

A runtime can declare what it is able to do. Only a recorded human decision makes an action authorized.

  1. 01Agent

    A coding agent proposes an action.

  2. 02Task

    The request is bound to defined work.

  3. 03Authority

    Identity is not treated as permission.

  4. 04Policy

    Kizma classifies the requested action.

  5. 05Approval

    Human authority stays in the loop.

  6. 06Evidence

    The decision and reported outcome are recorded.

A decision, not another dashboard

The requested action meets its control boundary.

Kizma makes the classification, the authority, and the required human action legible before a controlled runtime can proceed.

Claude Code
command_execution

Control decision

Human approval required

Runtime must stop before the action.
Task
Secure payment webhook
Repository
checkout-api
Classification
Repository-defined test run npm-test
Reason code
repository_defined_execution_requires_approval
Reason
Repository-defined execution is not automatically authorized, and Kizma does not attest the execution environment.
Evidence
Decision recorded; execution not authorized.
AWAITING HUMAN DECISIONAppend-only control record

Runtime neutral

One control model across coding agents.

Built to sit above the tools your engineering team already uses. Adapters translate runtime events into a single control contract without turning vendor identity into authority.

  • Claude Code
  • Codex
  • Cursor
  • Other coding agents

Kizma control contract

One decision vocabulary. One authority model. One record.

agent-control-contract-v1

Ecosystem names describe intended compatibility targets, not partnerships or endorsements.

Governance with provenance

Know what was requested, decided, and reported.

Durable action requests and append-only control events keep authorization separate from runtime-reported outcomes. Reported execution never becomes proof of authority.

  1. 01Request receivedaction_requestedTask-bound and classified by Kizma.
  2. 02Decision recordedrequire_approval / blockBlock or require human approval.
  3. 03Outcome reportedreported_executedPreserved without laundering authority.

Authorization and runtime-reported outcome stay separate. Version 1 records a block or a required human approval and issues no automatic allow; compliance is derived from the record, never declared by a runtime.

Design partner program

Let the agents do the work.
Keep the authority.

We're working with engineering teams that want agent speed without surrendering policy, approval, or traceability.

Become a design partnerExisting partner sign in
Kizma - AI Software Teams. Controlled.