better-chatbot 发布自托管 AI 聊天与智能体平台

An open-source AI chat and agent platform you can host yourself.

Talk to any model, attach your files, connect MCP tool servers, build workflows, and give agents a place to work — on infrastructure you control.

Deploy to Vercel

Two required values, a few optional ones. The button clones the repository, provisions a Neon Postgres database, and asks for:

The optional fields are pre-filled with optional; leave them like that and the feature stays off. You can add any key later under Settings → Environment Variables and redeploy. The schema is created during the build. Open the deployment and create an account; the first one becomes the administrator.

Details and optional features: docs/self-hosting.md.

Quick start (Docker)

Three variables. That is the whole configuration to get chatting:

git clone https://github.com/keinsaasforever/better-chatbot.git
cd better-chatbot
cp .env.example .env

POSTGRES_URL=postgres://postgres:postgres@postgres:5432/better_chatbot
BETTER_AUTH_SECRET=   # npx @better-auth/cli@latest secret
AI_GATEWAY_API_KEY=   # https://vercel.com/ai-gateway

docker compose -f docker/compose.yml up -d --build

Open http://localhost:3000 and create an account. The first account becomes the administrator.

Why one model key? Models are served through the Vercel AI Gateway, which fronts Anthropic, OpenAI, Google, Mistral, xAI, DeepSeek, Moonshot and more. You manage one key and one bill instead of seven, and you can switch models mid-conversation without touching configuration. If you would rather not use a gateway at all, point the app at Ollama, LM Studio, or any OpenAI-compatible endpoint — see Models.

Prefer running it directly? See docs/self-hosting.md.

What you get

Chat that does things. Streaming conversations with tool calling, file attachments, image generation, and a sandboxed code runner. Mention an agent, workflow, or tool with @ to pull it into the conversation.

Agents. Give an agent a role, instructions, and a set of tools. Share it with your workspace or keep it private.

Skills. Reusable capabilities an agent loads on demand, so a large toolbox does not have to sit in every prompt.

Projects with retrieval. Group chats, documents, and tools around a piece of work. Uploaded documents are chunked, embedded, and searched with pgvector — no external service required. Point a project at a knowledge base if you want hosted retrieval instead.

Workflows. Build a graph of steps — LLM calls, conditions, HTTP requests, code, templates — in a visual editor, publish it as a tool, and call it from any chat.

MCP tool servers. Connect any Model Context Protocol server over HTTP or stdio, with OAuth support, per-tool customization and a cached tool list. Self-hosted instances can run stdio servers; the desktop app runs them locally.

Scheduled tasks. Have an agent run on a schedule and report back.

Voice. Realtime voice conversations through the gateway, plus transcription.

Desktop app. An Electron build that adds local MCP servers, filesystem access, and the system keychain.

Model routing. Pick a model per message, or let the router choose from the catalog based on the task.

Upgrading from 1.x

This is a new major version. The layout is now a monorepo, the environment is different, and the schema has moved on considerably.

Back up your database first. Then:

pnpm install
pnpm tsx scripts/upgrade/from-better-chatbot-1.ts --dry-run
pnpm tsx scripts/upgrade/from-better-chatbot-1.ts --yes

The upgrade preserves your users, chats, agents, MCP servers and workflows. Archives become projects, and the threads inside them stay linked. It is a one-way migration — this is why the backup matters.

If you were setting per-provider API keys (OPENAI_API_KEY, ANTHROPIC_API_KEY and friends), replace them with a single AI_GATEWAY_API_KEY, or configure a local provider.

Setup

Vercel

The Deploy button above. Setting up an existing clone by hand is described in docs/self-hosting.md.

Docker Compose

The quick start above. The bundled stack includes Postgres with pgvector. Add --profile redis if you want resumable chat streams.

Local development

pnpm install            # writes .env from .env.example
pnpm docker:pg          # Postgres with pgvector on :5432
pnpm db:migrate         # apply the schema
pnpm dev                # http://localhost:3000

Postgres must have the pgvector extension — document embeddings are stored as vector(1536). The bundled image has it; managed Postgres usually offers it as a toggle.

Useful commands:

Optional integrations

Everything below is off until you configure it. The server prints which features are disabled at startup, so you always know why something is missing.

Running without Inngest. Scheduled tasks and the sandbox reaper are plain HTTP endpoints. Set SCHEDULED_TASK_SECRET and call them from any cron:

curl -X POST https://your-instance/api/scheduled-tasks/cron \
  -H "X-Scheduled-Task-Auth: $SCHEDULED_TASK_SECRET"

Models

The gateway is the default and covers the whole catalog. You can also:

• Run models locally. Connect Ollama, LM Studio, or any OpenAI-compatible endpoint from the model picker. Nothing leaves your machine, and it needs no server configuration — each account sets up its own.
• Use TensorX. EU-hosted inference with zero data retention, reached directly rather than through the gateway. Set TENSORX_API_KEY.
• Add any OpenAI-compatible provider with OPENAI_COMPATIBLE_DATA: pnpm openai-compatiable:init && pnpm openai-compatiable:parse.

Partners

TensorX — EU-hosted model inference with zero data retention. better-chatbot talks to TensorX directly rather than through a gateway, so prompts go straight to TensorX's EU infrastructure. Create a key at tensorx.ai, set TENSORX_API_KEY, and its models appear in the model picker next to the gateway's. See its trust and security page and DPA.

TensorX is optional. Without the key everything else works as before.

Community

• Discord — questions, help, and what people are building
• Issues — bugs and feature requests
• CONTRIBUTING.md — how to work in this repository

This edition, and the hosted one

better-chatbot is the community edition of keinsaas Navigator. Everything above is here, in full, with no usage caps and no telemetry.

What Navigator adds is what teams need and single users do not: shared team workspaces with member management and invitations, per-member AI budgets and org-wide model policy, billing, a managed workflow builder, and a Slack bot. If you are running this for yourself or a small group, you are not missing anything — that is the point.

License

MIT. See LICENSE.

Originally forked from cgoinglove/better-chatbot, whose work this is built on.