WhatIff
An open-source, self-hostable AI assistant where your continuity and memory are not tied to a single model provider.
WhatIff keeps the persistent AI state in your stack: conversations, long-term memory, personalities, tools, and scheduled agents. The model that generates the next response is interchangeable, so you can move between supported providers without making any one of them the owner of your AI's continuity.
WhatIff is the open-source core that powers whatiff.chat, a Go API server and an Angular web app you can run yourself. It owns the continuity layer independently of the model provider: semantic memory, personalities, conversation state, tools, and recurring work stay with the WhatIff deployment while supported models can be selected per user and per chat. Bring your own model API keys and host the whole thing on your own infrastructure.
Want the story behind it? Read the Imagination Foundry blog.
✨ Features
🤖 Multi-provider chat
- Chat with models from OpenAI, Anthropic, Google (Gemini), Z.ai (GLM), Mistral, DeepSeek, Qwen, and Xiaomi — configurable per user and per chat
- Streaming responses with Markdown rendering and an emoji picker
- Multiple chat sessions with sidebar navigation, inline renaming, and per-chat settings
🧠 Semantic memory
- Long-term memory stored as vector embeddings (pgvector) and recalled by conversation context
- Filter and search memories by date, scope, chat, and content
- Memory merging and compaction with checkpoints so context stays coherent over time
🎭 Personalities
- Define reusable AI personas with custom system prompts and behavior
- Switch personalities per chat or set a default; attach reference documents to a persona
- Full create/edit/organize UI
🔮 Rituals
- Reusable prompt templates you can invoke with customizable keyboard shortcuts
- Bind a ritual to a personality and auto-enrich messages with its context
🛠️ Tools & agents
- File attachments with 30+ supported types, plus semantic search over uploaded documents
- Code interpreter, image generation, and web search available to the assistant
- Agents — scheduled and background AI jobs with a distributed-lock-aware scheduler
- AI-generated files saved straight back into the conversation
🔗 Integrations
- Remote MCP connectors with per-chat activation controls
- Webhook API tokens (stored as one-way hashes) with static bearer auth isolated from session JWTs, each limited to the scopes it was created with (
messages:write, chat:read)
- Webhook trigger modes for existing chats:
user, assistant (write-only), and background (async agent trigger)
- Webhook read routes (scope
chat:read): threads (filterable by persona), persona names, messages, and job status; see the webhooks guide and the integration reference
🔧 Built for operators
- OpenAPI specification and structured logging (Zap)
- Ent ORM with automatic, non-destructive database migrations
- Docker & Docker Compose for local development
- GitHub Actions CI
🏗️ Tech stack
| Layer | Technologies |
|---|
| Backend | Go 1.27 · Gorilla Mux · Ent ORM · PostgreSQL 16 + pgvector · JWT auth · Zap |
| Frontend | Angular 22 · TypeScript 5 · Tailwind CSS 4 · RxJS · ngx-markdown |
| AI | OpenAI, Anthropic, Google, Z.ai, Mistral, DeepSeek, Qwen, Xiaomi (bring your own keys) |
| Infra | Docker & Docker Compose · GitHub Actions |
🚀 Quick start (Docker)
The fastest way to run the full stack locally is Docker Compose.
Prerequisites: Docker, and an OpenAI API key, which you add in the app.
# Start the stack. On first run it generates this install's secrets and runs
# migrations.
docker compose up --build
Then open http://localhost:8081, create an account, and add your OpenAI key when the
app asks (it lives in Profile & Settings → Providers, where you can add other
providers too: OpenRouter, a model server on this computer, or any compatible server).
Keys belong to your account and are stored encrypted; there is nothing to put in .env.
The JWT and token-encryption secrets are generated once per install and kept in a
Docker volume; nothing is shipped in the repository. Values you set in your shell or
.env take precedence. docker compose down -v deletes them along with the database.
Models on this computer. The API runs in a container, where localhost is the
container itself, so it reaches a model server on your machine (Ollama, LM Studio) as
host.docker.internal. Two things on the host have to allow that:
This is the beginning of the README. Read the rest at the original. Read the full README ↗
