Security

Own your AI team

Your server, your data, your agents and your tools. Koshra is self-hosted, built on open source and runs on free open models, and you are not locked to one model. Nothing is rented back to you.

• Your server. Koshra runs as Docker containers on a server you control, around the clock, not in someone else's cloud.
• Your data. Chats, tasks, the Brain and bot memory stay in the Postgres database on that server.
• Built on open source. Koshra is built on an open source agent platform released under the Apache 2.0 licence.
• Free open models. Bots can run on free open-weight models, such as NVIDIA Nemotron 3 Ultra and dots3-note, with no model fees. Our own Koshra team runs on them today.
• Not locked to one model. Pick the model per bot and switch mid-chat, or bring your own key for OpenAI, Anthropic, Google, OpenRouter, MiniMax or any OpenAI-compatible endpoint.
• Your agents and tools. Your bots, their jobs, their MCP servers and app connections live in your Koshra, under your own Allow, Ask and Block rules.

Model calls go to the provider you pick for each bot, and that provider sees the prompts it answers. The free open models are reached through OpenRouter.

Approval policy

Every tool call is checked against Allow, Ask or Block rules before it runs, for your account and for each space.

• Outbound Slack posts and emails wait for Allow or Deny on an approval card.
• Creating a bot, and creating or changing a schedule, need approval.
• Deleting, merging, purging or wiping CRM records is blocked.
• Bulk CRM changes, and any write over 25 records, need approval.
• Per-bot permission modes can make a bot stricter, never looser.
• Each dispatch has caps on spend, tokens, handoffs and runs, plus a loop guard.

Approvals and permission modes in the docs

Self-hosting

Koshra runs as Docker containers (web, API, worker, supervisor and Postgres) on a server you control. Protected pages and API routes refuse requests without a signed-in session. Credentials sit in the app's encrypted secret store.

Self-hosting in the docs

Data handling

• Chats, tasks, the Brain and bot memory are stored in the Postgres database on the Koshra server.
• Model calls go to the model provider chosen for each bot, so that provider sees the prompts it answers.
• Provider keys you add are stored encrypted and shown masked.
• The waitlist form on this site stores what you type in a database on our own server. It is not sent to any third party.

Operations

We take a backup and write a rollback note before every production change.

Certifications Not yet

Koshra does not hold a third-party security certification today. If you need a security review, talk to us.