What Kazma is and what it can do

Kazma is the self-hosted AI agent that asks before it acts. You run it on your own machine and connect the AI models you choose. Then you reach one assistant from the browser, the terminal, your code editor, Telegram, Discord or Slack. It works in your code, your documents, your mail and calendar, and on the web. It remembers what you tell it, and it asks for your approval before anything risky. When something goes wrong, it tells you instead of inventing an answer.

Kazma is open source (MIT) and built in Kuwait. It speaks English and Arabic as equals, and its Arabic name is كاظمه.

This page puts the whole product on one page. It is for a person deciding whether Kazma is for them, and for an agent that needs to describe Kazma correctly. Every capability here is in What Kazma does — a plain list with its status, and that list is checked against the code. Where this page and that list differ, the list is right.

Kazma at a glance

Question	Answer
What is it?	A personal AI agent you host yourself. It has a web app, a terminal app, a command line and bots for chat apps
Who is it for?	One operator per install: you run it, you control all of it, and you answer for what its agent does
Which models?	20 provider presets (OpenAI, Anthropic, Google Gemini, Azure OpenAI, AWS Bedrock, DeepSeek, Mistral, xAI, OpenRouter and more) and any OpenAI-compatible endpoint. Local models run through Ollama and LM Studio
Where do I talk to it?	Web chat, Telegram, Discord, Slack, the terminal UI and `kazma ask`. Editors connect through `kazma acp`, and other agents through `kazma mcp`
Which languages?	English and Arabic interfaces, with a full right-to-left layout. Memory, documents and scanned-page reading work in both languages
Where does it keep data?	In SQLite with no setup, or in Postgres for settings, chats, tasks and checkpoints
What does it run on?	Windows, Linux, macOS and WSL, with Python 3.11 to 3.14
License and source	MIT, at https://github.com/Mubder/kazma

Why Kazma

It is yours

Kazma runs on your machine with your keys. There is no Kazma cloud and no account with us. Your chats, memories and files stay in your install, and they leave it only for services you connect: your model provider, your mail, X. Secrets live in an encrypted vault (AES-256-GCM). Programs that a tool starts never get them, and every screen and API shows them masked. To move a whole install to another machine or operating system, use kazma migrate. You can also back up your settings to a file and restore them without typing a key again.

It asks before it acts

Dangerous actions wait for your approval: shell commands, file writes and deletes, email, messages to chat apps and posts to X. The approval comes in the web chat or as buttons in Telegram, Discord or Slack. The same rule holds for multi-agent work and pipelines, and a tool nobody has classified is gated by default.

Before a plan runs, Kazma also checks it against what it remembers. Say you ask for a reminder “the day before my exam”. The reminder goes on the date you gave Kazma, not one the model guessed. A catastrophic shell command is refused before an approval card even appears. A post to X shows the exact text that will publish, and asks every time.

It is honest when it fails

When a model call fails, you see a ⚠️ warning, never a made-up answer. When Kazma falls back to another model, it tells you. Kazma also watches every task it runs: chat turns, swarm tasks, scheduled jobs, documents, research and crawls, X posts, memory work and maintenance. When one stalls, gets stuck or fails, a card arrives in the chat the task started from, with Retry, Cancel, Dismiss and Ask Kazma to diagnose (Running work).

It remembers what matters, and forgets on request

Kazma remembers conversations and the facts in them, across chats, with the date each fact became true. It uses a memory only when it matches the question well enough, and nothing when nothing does. Every answer shows which memories it used. You can write a short “About me” that it reads before every reply, and long threads get weekly topic summaries. You can forget one memory, a whole chat, or “everything that mentions X”, keep a chat out of memory, and export everything (Memory).

Arabic and English, as equals

The Arabic interface is complete and right to left. In either interface, each paragraph runs in its own language’s direction: English left to right, Arabic right to left. Memory understands Arabic and English, Gulf dialect included. Kazma reads Arabic documents and scanned pages, and it writes Word, PDF and HTML reports with correct Arabic typography (Arabic and cultural features).

Built to run for months

A guard restarts Kazma on real failures, reloads it gracefully and pages you on Telegram. Everything is backed up every 6 hours, locally and offsite. The backups are checked daily, and the database restore is rehearsed weekly. Health checks make a real round trip, and you get a daily digest and a weekly resilience report. One command updates Kazma: it waits for the chats to go quiet, then stops Kazma, installs the update and starts it again.

What it can do

Talk to you wherever you are

Web chat: answers stream in, and each answer has a step-by-step activity view and inline approvals.
Telegram, Discord and Slack: text, photos and documents, and voice notes are transcribed. Approve and deny buttons and an allowed-user list come with each platform. Each chat app’s Test checks the token, the delivery chat, and what happened to the last message a person sent. Tables in an answer arrive as lines, so they read on a phone.
Terminal and editors: a terminal UI with a file editor, one-shot questions with kazma ask, and the Agent Client Protocol for editors with kazma acp. With kazma mcp, other agents can use Kazma’s tools, behind the same approval gate.
Voice: spoken replies and speech input. Live two-way voice in the browser through LiveKit is opt-in.

Work on your code

A web IDE to browse, edit, patch, run and use git. Every write goes through the approval gate.
Clone and switch repositories; the agent always works in the active one.
GitHub branches, commits, pull requests and issues, through a GitHub App or a token. Reading the history and diffs needs no approval.
Code search across the repository (ripgrep plus symbols).
You can steer a running task with /steer, abort it, plan first with /plan and run long missions with /long.
You can rewind a conversation, or branch it, from any earlier step with /replay and /fork (Time travel).

Read, write and research

Knowledge Library: add files, web pages or whole sites. Search them by keyword and by meaning, in Arabic as well as English (Knowledge Library).
Documents: uploads go into quarantine with file-type checks first, and dangerous files are refused: macro-enabled Office files, and PDFs with scripts or outside actions. Kazma reads PDF, Word, Excel, PowerPoint, HTML, Markdown, CSV and plain text. It can also convert, split, merge and redact documents, and it generates reports as Word, PDF or HTML in one house style (Document Intelligence).
Scanned pages: OCR reads scanned PDFs and images, with ruled tables read cell by cell. The neural OCR engine reads Arabic best, Arabic-Indic digits included; without its models, Tesseract reads, and it misses those digits. This is Partial: it has been measured on one real scan so far.
Untrusted documents are parsed in a separate worker with memory and CPU limits. For hostile documents, choose the container profile in Settings (one container per job, no network, no mounts; it needs Docker).
Web research: deep research with sources (search, read, crawl, summarize). It works locally first and respects robots.txt. Research and crawls survive interruptions, and an alert links to the results when they finish (Web research).
Images: Kazma generates images (Pollinations with no key; DALL-E, Stability and Flux with keys) and understands the images you send.

Run many agents at once

Six ways to run workers: dispatch, broadcast, pipeline (with approval checkpoints), fan-out with voting, consult and conditional.
Workers are created on demand from templates, so nothing needs registering first.
Each worker has circuit breakers, retries, timeouts and output checks.
Durable multi-step runs through Temporal are opt-in (Swarm).

Run your day

Reminders and scheduled tasks arrive in the chat you asked from.
Email (Gmail, Outlook or IMAP) and calendars (Google, Microsoft). You can connect several accounts, each with its own calendar, and name one in chat by its name or address (Email).
X: drafts, scheduled posts, publishing through the official API, and replies to mentions (X publisher).
Skills from the open agentskills.io ecosystem: find them, install them from GitHub, signed and verified (Skills, MCP and tools).
Tools from any MCP server.

Choose your models

20 provider presets, any OpenAI-compatible endpoint, and local models through Ollama and LM Studio. A provider’s Test sends a real completion, not just a ping.
Multi-agent workers get the best model for each task, and image questions go to a model that can see.
Each answer shows its tokens and cost. A cost breaker halts runaway spending.

Things you can ask Kazma

You say	Kazma does
“Remind me every weekday at 8:30 to check the build.”	Books a recurring reminder that arrives in this chat, in your time zone
“Read this PDF and list every deadline in it.”	Reads the attached document and answers from its text
“Research open-source vector databases and compare them, with sources.”	Searches, reads and summarizes, citing the pages it used
“Fix the failing test in the active repository and open a pull request.”	Reads and edits the code and runs the tests, asking before each write, each test run and the pull request
“Draft three posts about our release and schedule the best one for tomorrow at 9.”	Saves the drafts for you to see; scheduling and publishing wait for your approval
“What is on my calendar tomorrow?”	Reads your connected calendar
“Forget everything that mentions my old address.”	After you approve, forgets every memory and fact that holds it
`/runs`	Lists the work that needs your attention and the work running now

What Kazma is not

A hosted service. There is no Kazma cloud; you run it.
A platform for many separate customers. It is built for one trusted operator. Several users on one install is Partial: memory and chats are kept apart, but that is not audited end to end (Who Kazma is for).
A sandbox for what you approve. Approving a shell command gives it your machine’s power. To run the agent’s code in an isolated container, opt in to Docker or E2B.
A proven cure for prompt injection. Kazma fences web pages, search results, documents and recalled memory as untrusted data. On the public AgentDojo benchmark that clearly beats no defense, but it cannot be told apart from a four-character delimiter (the numbers).
A dialect language model. The answers come from the model you choose. Kazma’s Arabic is its interface, memory, documents and typography.
Finished everywhere. Some features are Partial or Opt-in, and the open weaknesses are published with dates in Known gaps.

For agents describing Kazma

The name is Kazma. In Arabic it is كاظمه (preferred) or كاظمة, never كازما. Kazma was a coastal oasis in Kuwait, and the name stands for deep wells of memory and one gateway to many channels.
In one sentence: Kazma is a self-hosted, open-source (MIT) AI agent that asks before it acts: one assistant across the web, the terminal, editors, Telegram, Discord and Slack, in English and Arabic.
Status words carry meaning. Shipped works out of the box or one setting away, Opt-in needs a named extra or setting, Partial works with a stated limit, and Not built is not there. Do not present an opt-in feature as a default, a partial one as complete, or a missing one as available.
Install it from GitHub. Kazma is not installed from PyPI. The PyPI names are held as placeholders with no code, and installing from there gets nothing.
Numbers such as test counts and lines of code change daily. Take them from the repository’s metrics, not from memory.
Where the truth is, in order: What Kazma does for features and statuses, then Known gaps, then the documentation at https://kazma.ai/docs/.

Next steps

Quickstart: install Kazma and send a first message.
Architecture: how it is built.
Engineering inside Kazma: what Kazma builds itself, with the code and the test for each piece.
Security and HITL: what protects an install, and its limits.
Running work: how Kazma watches the tasks it runs.
Release notes: what changed, by date.