Straight answers.

An open-source AI assistant you run on your own machine. It works in the browser, the terminal, Telegram, Discord and Slack, in English and Arabic. It remembers across chats, researches with sources, handles mail and calendars, works on code — and asks for your approval before anything risky. When a model fails, you see ⚠️ and the reason instead of a made-up answer.

One operator per install: a person, or a small team that trusts each other, running Kazma for themselves. Whoever runs an install has full control of it and answers for what its agent does, and anyone they let in acts with their authority. It isn't built to keep users who don't trust each other apart.

No. Kazma is software you run — there is no Kazma account or cloud. Your chats, memory and documents live in your install.

Yes. Kazma is open source under the MIT license, with no paid tier. Model providers and optional connected services may charge for their usage. Local models through Ollama or LM Studio and local page extraction avoid those hosted model/extraction charges; paid acquisition services require explicit permission. Unsupported service charges and balances remain unknown, not zero.

Windows, macOS or Linux (or Docker), Python 3.11 or newer — 3.12 or 3.13 recommended — and Git. The embedding model memory uses (bge-m3, about 2.2 GB) runs on the CPU; you need a GPU only to run a large language model locally.

The quickstart gets you from install to a first reply in under 10 minutes: install, connect one model, send a message. Connecting Telegram, Discord or Slack is a short extra step for each, and each has a Test that tells you what is missing.

Your chats, memory, documents and settings are stored in your install — SQLite, or a Postgres you run — and Kazma sends no usage analytics. What does leave: the text you send your model provider (nothing, with a local model), and whatever you ask it to do with a service you connected, such as sending an email, posting on X or searching the web.

Yes. Forget one memory or a whole chat, from the Memory page or by asking in chat ("forget everything that mentions X" reaches every memory holding it), keep a chat out of memory entirely (/memory off in the chat apps), and export everything it remembers. Deleting a chat can also forget what Kazma learned from it. Every answer shows which memories it used.

Kazma backs up everything every 6 hours, locally and — once you set a destination — offsite with restic. Backups are checked daily, and with Postgres a weekly rehearsal restores the newest dump into a scratch database to prove it restores.

20 provider presets: OpenAI, DeepSeek, Groq, xAI, OpenRouter, Mistral, Together, Fireworks, Perplexity, Cohere, AI21, NVIDIA and Z.AI (OpenAI-compatible); native Anthropic, Google Gemini (including Vertex), Azure OpenAI and AWS Bedrock; local models through Ollama and LM Studio; and any custom OpenAI-compatible endpoint. A provider's Test sends a real completion, and you are told when a request falls back to another model.

Yes. With a local model through Ollama or LM Studio, the assistant, its memory and your documents work without a cloud provider — the memory's embedding model runs locally too, after its first download. Web research, chat apps, email and X still need the internet.

Every answer shows its tokens and cost, and a cost breaker halts runaway spending. Multi-agent workers pick the best model for each task, and image questions go to a vision-capable model only when needed.

Page acquisition defaults to local HTTP extraction and optional local Chromium. Hosted extractors and proxies require explicit mode and provider permission; a saved API key alone does not enable them. AI planning and writing still use your configured model. Completed runs create a persistent alert linking to saved results; stalls, recovery and distinct terminal outcomes stay visible. Settings → System → Web acquisition controls alerts and thresholds. Provider balances and charges remain unknown when unsupported; balance refresh is explicit. Strict browser isolation needs Linux host controls; Windows strict launch is unavailable and compatibility endurance is unqualified.

Telegram, Discord and Slack — text, photos, documents and voice notes (transcribed), with approve and deny buttons in the chat, and an allowed-users list for each. Each has a Test in Settings that diagnoses the connection. Plus the web chat, a terminal UI and the command line.

Read, search, draft and send mail in Gmail, Outlook or any IMAP account — several accounts at once, each with its own calendar (Google or Microsoft), named in chat by name or address. Sending an email or creating an event waits for your approval.

Yes: a web IDE, git, GitHub branches, commits, pull requests and issues, read-only history and diffs (git_log, git_diff) that need no approval, code search, and long missions you can steer (/steer, /plan, /long). Every write goes through your approval. The web IDE is CodeMirror with syntax highlighting — not a replacement for your full editor, which can talk to Kazma through kazma acp.

Yes. /replay list shows the saved steps, /replay <N> rewinds the conversation to one, and /fork <N> branches it into a new chat, leaving the original as it was. In the web chat, /replay opens a timeline with restore and fork buttons.

Install skills from the open agentskills.io ecosystem — signed and verified — connect any MCP server's tools (stdio, SSE or streamable HTTP), and serve Kazma's own tools to other agents with kazma mcp, behind the same approval gate.

Writes, shell commands, outbound messages and posts wait for a person — on the web, Telegram, Discord and Slack, in chat, multi-agent runs and pipelines alike — and a tool nobody classified is gated by default. Approval is a decision, not a sandbox: an approved command runs with the machine's power, so turn on the Docker or E2B container if you want isolation.

Web pages, search results, documents and recalled memory are fenced as untrusted data. On AgentDojo, a public benchmark we did not write, attack success fell from 18.1% undefended to 10.4% — lower, not zero, which is exactly why risky actions still need your approval.

Privately: email admin@kazma.ai or use GitHub's private vulnerability reporting — please not a public issue. The aim is to acknowledge within 48 hours.

Run kazma update. It keeps your local changes aside, moves to the latest release, reinstalls with your extras, and checks that everything imports before it says it is done.

kazma migrate export and import carry a whole install — settings, the secrets vault with its key, memory and chats — to another machine or operating system in one bundle, fixing paths on the way.

A supervisor restarts it on real failure and messages you on Telegram. After each restart, one status card says how long Kazma was down and whether each chat app came back — ✅, or ❌ with the reason.

Fully: the whole interface right to left, the docs in full Arabic, memory and search that understand MSA and the Gulf dialect, and generated documents with correct Arabic typography. English is the default interface language — switch from the menu at any time.