Skip to content
Kazma Docs
GitHub

Welcome to Kazma

Kazma (كاظمه) is a multi-platform, autonomous AI agent framework built on a LangGraph supervisor brain, swarm orchestration, cross-platform dispatch (Telegram / Discord / Slack / Web / TUI), and an OpenAI-compatible LLM provider layer — with first-class Arabic / RTL support and a Gulf-cultural conversational protocol (Majlis).

These docs are a ground-up, code-verified reference: every factual claim is traceable to source, and anything not yet implemented is explicitly marked.

Quickstart

Install, configure one provider, and send your first message in under 10 minutes.

Architecture

The supervisor ReAct loop, data-flow diagrams, and subsystem internals.

Configuration

Every kazma.yaml key, every env var, and the ConfigStore override model.

Security & Safety

The three HITL gates, danger-tool lists, and fail-closed behavior.

Where to go next

Section titled “Where to go next”

The legacy of the name

Section titled “The legacy of the name”

Kazma (كاظمة) was an ancient coastal region in Kuwait — famed as a vital oasis network of water wells and a strategic trade gateway connecting distinct civilizations. It was also the site of the legendary Battle of Chains (Dhat al-Salasil, 633 CE), where strict, unified coordination broke a rigid defensive line through precise, decentralized maneuvering.

The framework inherits these three structural principles — described honestly, as the code implements them today:

  • The Wells → Memory. Long-term memory that holds context across sessions and from which the agent draws when it needs to remember. (Memory retrieval is opt-in via the memory_search tool — see Memory & RAG for the real wiring.)
  • The Gateway → Routing. A single LangGraph supervisor routed out to many channels — Telegram, Discord, Slack, Web, and the TUI — through one adapter layer. (See Gateways & Platforms.)
  • The Chains → Trust. Every skill is HMAC-signed and every dangerous action passes through three independent human-in-the-loop gates. (See Security & Safety.)

Documentation honesty: Kazma distinguishes what the code does today from what older copy claims. Features that are only partially wired are flagged inline under “Honest status notes” — most notably in Memory & RAG and Troubleshooting.