kai
Get started
Kai

AI chat for your team,
on your own servers

Kai is an open-source chat interface. Bring your own keys, connect MCPs, and keep every conversation in your own server.

Get started Star on GitHub

Free and MIT licensed.

A conversation in Kai: Claude answers with a table comparing Ada Lovelace’s Note G with a modern for-loop
Works withClaudeGPTLiteLLMBifrostAzure OpenAI130+ MCP servers

From a question to work that’s done

People ask in a shared workspace. Kai reaches your models and tools, checks before anything changes, and keeps it cheap.

How Kai works

1Bring your own keys

Connect your LLM providers and MCP servers in one place.

AdaOwnerGraceAdminAlanMemberInferenceOpenAIAnthropicGoogleDeepSeekAzureGatewaysLiteLLMBifrostMCPsLinearNotionSlackGitHub

2Kai asks before it acts

Each tool call needs your approval.

File the Safari SSO bug in Linear, high priority.
Create issueLinearCalling

3Every step reads the cache

Each request starts with the one before, so long chats are billed mostly at the cached rate. Kai tries to maximize cache hit.

Cache hit ratio0%
from cache newIllustrative

4Create agents

Create agents with their own instructions, tools and model allowlist.

HHistorianEveryone in the workspaceDraft

Instructions

Tools

Researchtool groupNotesMCP server

Models allowed

claude-opus-5-5
claude-sonnet-5-5
gpt-6.1-sol

Every MCP server, with guardrails

Connect the tools your team already works in. Kai handles sign-in, permissions and approvals.

MCP servers in the docs

A library of 130+ servers

GitHub, Linear, Notion, Sentry, Stripe… Pick one and Kai fills in how to connect.

Every way MCP signs in

OAuth, headers or client credentials, shared by the workspace or per person.

Permissions per tool

Allow, needs approval or deny, for a whole server, a group or a single tool.

Any HTTP API as tools

Point Kai at an OpenAPI document and each operation becomes a tool.

Agents and tool groups

Give a team its own instructions, models and tools; switch sets on together.

Usage you can read

Tokens and cache hit ratio by day, member and model. No guessed prices.

GitHubLinearNotionSentryStripeFigmaAtlassianAsanaAirtableCloudflareVercelSupabasePostHogIntercomHubSpotSlackShopifyBoxMondayNeonMiroHugging FacePlaywrightCanva

Calm, fast and familiar

Streaming that survives a refresh, branches for edits, attachments, web search and reasoning effort, in light and dark.

A conversation in Kai, in light mode, with a table in the answer

Any model, on your keys

Each workspace adds its own keys, stored encrypted and never shown again. Use a provider directly or the gateway you already run to reach any model, and switch models mid-chat.

AnthropicMessages APIClaude with adaptive thinking and its signatures replayed, and explicit cache points.
OpenAIResponses APIGPT with encrypted reasoning carried across steps and a cache key per chat.
LiteLLMGatewayPoint a key at the proxy and reach every model it routes, with its virtual keys.
BifrostGatewayNative Anthropic and OpenAI paths, so nothing is translated away.
Azure OpenAIOpenAI v1Your deployments, by name, through Azure’s v1 API.
Anything compatibleMessages or ResponsesIf it speaks one of the two APIs with streaming, Kai can use it.

Up in one command, on your servers

Fill in a few values in .env and start the stack. Migrations run on their own, on every upgrade.

  • Bring your own Postgres, Redis and S3, or use the ones in the Compose file.
  • Run several replicas behind your proxy; a reply streams through any of them.
  • Keys and tokens are encrypted with a key only you hold.
Self-hosting guide
~/kai
$

Give your team AI it can trust

Open source under the MIT license. Run it today; there’s nothing to sign up for.

Read the quickstart Star on GitHub