Frequently asked.

Linux, macOS, Windows, Docker, and WSL. Deploy with a single `docker compose up` — 2 volumes, graceful shutdown, no OS-specific code.

All user state lives under `~/.kazma/`; databases in `kazma-data/`. Every path is overridable via config or env vars — no hardcoded paths.

Yes — Kazma snapshots the full agent state after every supervisor iteration. Use /replay list to see snapshots, /replay <N> to rewind the live thread to that point, /fork <N> to branch into a new thread (original stays intact), /replay compare <A> <B> to diff two snapshots, or the /replay Web UI panel for a visual timeline with restore and fork buttons. Snapshots persist in kazma-data/snapshots.db.

An open-source framework for building reliable, culturally-aware AI agents. Built on LangGraph with SQLite or optional Postgres checkpointing, it survives crashes, remembers across sessions, and enforces safety boundaries — with a multi-platform gateway, RAG memory, and human-in-the-loop safety.

Clone the repo, run `uv sync`, then launch `kazma-web` (Web UI), `kazma-tui` (Terminal), or `kazma` (CLI). Windows users can run `.\setup.ps1`.

Yes — Kazma is built with native Arabic support. It handles MSA and Gulf dialects with a Majlis Mode protocol for Gulf cultural context, plus a full bilingual UI.

Eighteen presets ship out of the box via the OpenAI-compatible provider layer — OpenAI, Anthropic, DeepSeek, Google Gemini (ADC), Groq, xAI, OpenRouter, Mistral, Together, Cohere, Fireworks, Perplexity, AI21, NVIDIA NIM, plus local runtimes (Ollama, LM Studio). Anthropic, Azure OpenAI, and AWS Bedrock have dedicated native provider classes (they use their own auth/schema, not the OpenAI wire format), so no proxy is required. The `models.router` string gates a LiteLLM-proxy fallback branch, but LiteLLM itself is not imported.

Triple-wired human-in-the-loop gates intercept dangerous operations at the graph, swarm bus, and pipeline-checkpoint layers, arbitrated by one gate registry (hitl_gates.db). Danger tools require human approval on Web, Telegram, Discord, and Slack, and are fail-closed by default.

Six dispatch patterns — dispatch, broadcast, pipeline, fan-out, consult (independent opinions + synthesis), and conditional (router-based) — plus auto-routing (capability-based worker selection). Reliability layer: RetryPolicy, CircuitBreaker, OutputValidator.