# Capka

> Capka is a self-hosted agent platform that returns files, not chat replies. You attach spreadsheets, documents, archives or code and describe the result you want; it works inside an isolated Linux sandbox and returns a report, a cleaned dataset, a converted file or a patch.

Capka runs on your own Linux server with Docker — 2 GB RAM and 20 GB of disk are enough to start — and uses your own model API keys (Anthropic, OpenRouter, any OpenAI-compatible API) or local models through Ollama. Every chat gets its own Linux sandbox with Python, Node, LibreOffice, FFmpeg, Playwright and OCR, and tasks keep running on the server after the browser tab is closed.

One instance serves many people: files and chats stay private per user, while admins see roles and token spend rather than other people's conversations. The full platform is also available through Telegram. It is extended with Anthropic-compatible skills, MCP connectors and marketplace bundles.

There is no Capka cloud and no per-seat subscription. You run the server and pay your model provider directly. The software is open source under AGPL-3.0.

Install with:

```
curl -fsSL https://raw.githubusercontent.com/LyoSU/capka/master/install.sh | sh
```

## Docs

- [Landing page](https://capka.app/): what Capka does, who it suits, and how it compares to Claude Cowork, OpenClaw, Suna, Open WebUI and Onyx
- [Ukrainian landing page](https://capka.app/uk/): the same page in Ukrainian
- [Documentation](https://capka.app/docs/): technical overview and documentation index
- [Installation](https://capka.app/docs/installation/): requirements, one-command deployment, first run and updates
- [Configuration](https://capka.app/docs/configuration/): access, storage, model keys, policies and connectors
- [Environment variables](https://capka.app/docs/environment/): all production secrets, routing, sandbox, retention, worker, MCP and telemetry settings
- [Models](https://capka.app/docs/models/): Anthropic, OpenRouter, OpenAI-compatible endpoints and Ollama
- [Sandbox](https://capka.app/docs/sandbox/): per-session Linux containers, execution tools, limits and hardening
- [Files and tools](https://capka.app/docs/files/): persistent workspaces, Python, Node.js, LibreOffice, FFmpeg, OCR and Playwright
- [Security](https://capka.app/docs/security/): isolation, user privacy, permissions, egress, gVisor and rootless Docker
- [Administration](https://capka.app/docs/admin/): users, roles, model access, policies, audit events and usage
- [Troubleshooting](https://capka.app/docs/troubleshooting/): installation, model and sandbox diagnostics
- [Full documentation in one file](https://capka.app/llms-full.txt): concatenated Markdown for retrieval and offline context
- [Ukrainian summary](https://capka.app/uk/llms.txt): this file in Ukrainian
- [Source repository](https://github.com/LyoSU/capka): code, issues and releases
- [Installer script](https://github.com/LyoSU/capka/blob/master/install.sh): the exact script the one-line install runs
- [Security notes](https://github.com/LyoSU/capka/blob/master/SECURITY.md): sandbox isolation and hardening with gVisor or rootless Docker
- [License](https://github.com/LyoSU/capka/blob/master/LICENSE): AGPL-3.0

## Architecture and isolation

The deployed stack contains the Next.js platform and worker, PostgreSQL task queue, sandbox controller, restricted Docker socket proxy, and the sandbox execution image. Agent code runs in per-session containers with `no-new-privileges`, dropped Linux capabilities, resource limits and commands pinned to uid 1000. A session mounts only its user's workspace paths. Fresh installations have no sandbox egress until an administrator enables it; deployments can hard-disable egress or restrict it to an allowlist. For untrusted multi-tenant use, Capka documents gVisor plus rootless Docker as defense in depth.

## Optional

- [Demo instance](https://demo.capka.app/): hosted by the maker; requires an account