Your own AI assistant.It asks before it acts.
Run Kazma on your machine with your choice of cloud or local model. Research, write and work on your code in English or Arabic, from the web or your chat apps, with your approval before risky actions.
- Free and open source
- Any model, including local ones
- Your chats and memory stay in your install
Works where you already are
- Web
- Terminal
- Command line
- Telegram
- Discord
- Slack
- Your editor (ACP)
- Other agents (MCP)
An assistant that gets things done, not just a chat box.
Nine areas, all working today. The full list — every line with its status — is on the Features page.
Chat everywhere
One assistant in the browser, the terminal, Telegram, Discord and Slack — it reads photos, documents and voice notes, with approve and deny buttons right in the chat.
DetailsMemory you control
Remembers people, facts and decisions across chats, and when each became true. Every answer shows what it used; make it forget anything, or keep a chat out of memory.
DetailsAsks before it acts
Writes, shell commands, posts and messages wait for your approval on every path. Plans are checked against what you actually said, and catastrophic commands are refused outright.
DetailsAny model, even local
20 provider presets — OpenAI, Anthropic, Gemini, DeepSeek and more — or local models through Ollama and LM Studio. See the cost of every answer; a breaker stops runaway spending.
DetailsKnowledge and documents
Add files, pages or whole sites and search them by meaning. Upload PDF, Word, Excel or PowerPoint safely; convert, redact, and generate reports with correct Arabic typography.
DetailsResearch with sources
Research with saved sources and results, completion alerts and progress warnings. Local acquisition by default; hosted extractors need explicit consent. AI planning and writing use your chosen model.
DetailsMail, calendar and reminders
Gmail, Outlook and IMAP accounts with their calendars; reminders delivered to the chat you asked from; X drafts, scheduling and posting; skills and MCP tools.
DetailsCode and repositories
A web IDE, git and GitHub pull requests, code search and steerable long missions — every write with your approval. Rewind or fork any conversation.
DetailsMany agents at once
Six ways to split work across workers created on demand, with retries, circuit breakers and output checks per worker.
DetailsFrom install to first reply in minutes.
Three steps. No cloud account, no subscription.
Install it
One script on Windows, macOS or Linux, or Docker Compose. Needs Python 3.11 or newer.
Connect a model
Pick any of 20 providers or a local model. Its Test sends a real completion, so you know it works.
Connect your chat apps
Add Telegram, Discord or Slack. Each has a Test that tells you exactly what is missing, if anything.
# 1. Get the code$ git clone https://github.com/Mubder/kazma.git$ cd kazma# 2. Install (fetches uv if needed)$ .\setup.ps1# 3. Start — chat opens on http://localhost:9090$ kazma serve# 1. Get the code$ git clone https://github.com/Mubder/kazma.git$ cd kazma# 2. Install (fetches uv if needed)$ ./setup.sh# 3. Start — chat opens on http://localhost:9090$ kazma serve# 1. Get the code$ git clone https://github.com/Mubder/kazma.git$ cd kazma# 1. Create your settings file, then edit it$ cp .env.example .env# 2. Start — chat opens on http://localhost:9090$ docker compose up -d --buildSee it working.
Real captures of Kazma running on a machine, with demo data.
Every answer shows its steps: the tools it ran, what it remembered, and what it cost.
What Kazma remembers, in front of you: facts, conversations and weekly summaries — each one can be forgotten.
A web IDE with git and a chat that knows the open file. Every write goes through your approval.
Telegram, Discord and Slack in one place, each with a Test that diagnoses the connection.
The same interface in Arabic, right to left — switch the language from the menu at any time.
Trust you can check.
Every number here has a source, and the limits are published next to the features.
On AgentDojo, a public benchmark we did not write: 18.1% undefended, 10.4% with Kazma's fence. A 4-character delimiter from the literature scores 11.6%, and the two can't be told apart statistically — we say so.
The full numbersRun on every change. Fixes land with a test that reproduces the bug, and a check that keeps it from coming back in similar places.
Continuous integrationEverything is backed up locally and offsite, checked daily, and a weekly rehearsal restores the database into a scratch copy to prove it restores.
Disaster recoveryWhat Kazma doesn't do yet is written down, dated, in the docs — and the feature list gives every line its status: Shipped, Opt-in or Partial.
Known gapsWhen a model fails, you see ⚠️ and the reason — never a made-up answer.
Arabic, done properly.
Kazma is built in Kuwait, and its Arabic is not an afterthought.
A full right-to-left interface
Every page in English and Arabic, the login page included — switch at any time.
Search and memory that read Arabic
Search folds diacritics and hamza forms; memory understands MSA and the Gulf dialect.
Documents with correct Arabic typography
Word, PDF and HTML reports shaped correctly, with direction set per paragraph — even in a bilingual document.
Docs in both languages
Every documentation page is available in full Arabic, not as a summary.
Built to run for months.
It restarts itself, backs itself up, and tells you what happened — in the chat app you choose.
A supervisor watching it
Restarts on real failure, reloads gracefully, and pages you over Telegram.
One card per restart
How long it was down, and whether each chat app came back — ✅ or ❌ with the reason.
Backups that prove themselves
Every 6 hours, local and offsite, with a weekly restore rehearsal.
Updates in one command
kazma update pulls the new release and applies it.
Move it to any machine
kazma migrate carries a whole install to another machine or OS in one bundle.
Postgres or SQLite
SQLite with no setup, or Postgres; Prometheus metrics and OpenTelemetry traces.
Who is Kazma for?
One operator per install: whoever runs it has full control of it and answers for what its agent does. The safeguards protect the operator from the agent and the untrusted content it reads — not an install from its operator.
Run your own assistant today.
Open source under the MIT license. No account and no subscription — your data stays in your install, and goes only to the model and services you connect.
- automated tests
- 19,394
- public commits
- 4,467+
- packages
- 6
- contributors
- 10