"Built the day DeepSeek went down — and I had no Plan B."

Switch AI providers with one click. Never lose context when one goes offline.

It was a normal workday. DeepSeek open, ready to go — then: down. Completely gone. No fallback, no Plan B.

I had to manually copy everything into another interface. Context lost. Time lost. Flow lost.

One week later, 20centAI existed.

A minimalist Streamlit chat interface that:

• connects 8 AI providers — Claude, GPT-4o, Mistral, DeepSeek, Perplexity, Qwen, Groq and more
• auto-compresses conversation history — you pay 90% fewer tokens
• runs in a single Python file (~600 lines)
• needs no frameworks (no LangChain, no vector database)

Every message you send ships the entire conversation history along with it. After 50+ messages you're paying for tokens you don't need — and most chat tools do nothing about it.

90% saved. Automatically. Invisibly.

Messages 1–20  →  compressed into a single summary (~$0.001)
Messages 21–40 →  kept verbatim  ← the AI always sees fresh context

When the conversation hits 40 messages, the oldest 20 are summarized automatically by whichever model you're currently using. You keep full context. You pay 90% less.

Existing solutions are either buried inside large frameworks with significant setup overhead, or require separate ML models that are overkill for casual API users.

20centAI fills the gap: a single-file, framework-free implementation of rolling-window compression that anyone can read, fork, and modify in an afternoon.

Tested on Linux (Ubuntu 22.04+). Should also work on macOS and Windows WSL2.

# 1. Create and activate virtual environment
python3 -m venv .venv
source .venv/bin/activate

# 2. Install dependencies
pip install streamlit anthropic requests python-dotenv

# 3. Create .env file with your API keys
cat > .env << 'EOF'
ANTHROPIC_API_KEY=sk-ant-...
OPENAI_API_KEY=sk-...
MISTRAL_API_KEY=...
DEEPSEEK_API_KEY=sk-...
PERPLEXITY_API_KEY=pplx-...
DASHSCOPE_API_KEY=sk-...
GROQ_API_KEY=gsk_...
EOF

# 4. Run
streamlit run ai_chat_en.py
# Opens automatically in your browser at http://localhost:8501

# Next time just:
source .venv/bin/activate && streamlit run ai_chat_en.py

All keys are optional — unavailable providers are disabled automatically at startup. Add only the keys for the providers you actually use.

Adding a new OpenAI-compatible provider = 4 lines in the MODELS dict, zero code changes.

COMPRESS_THRESHOLD = 40  # trigger compression after N messages
KEEP_ORIGINAL      = 20  # always keep last N messages verbatim

Adjust to your workflow:

• Long research sessions → 60 / 30
• Quick daily driver → 30 / 15

Before (40 messages, ~8000 tokens sent per request):
  [msg 1] ... [msg 40]  ← full history every time

After compression (~1800 tokens):
  [📋 SUMMARY of msg 1–20]  ← ~300 tokens, generated once
  [msg 21] ... [msg 40]     ← last 20 verbatim

The summary is generated once by the active model, then reused in every subsequent request — paying ~$0.001 instead of ~$0.10 per round.

ai_chat_en.py       ← the entire app (English UI)
ai_chat_de.py       ← same app (German UI)
council_chat.md     ← active chat session (plain text, human-readable)
ai_council.db       ← SQLite archive with FTS5 index
.env                ← API keys

User Input
    │
    ├── @archive?  →  FTS5 search  →  inject context into prompt
    └── normal     →  ai_response()
                           │
                    Two-branch design:
                    ├── Claude       → Anthropic SDK
                    └── Everyone else → OpenAI-compatible REST
                           │
                    save to SQLite (append-only, immutable)
                           │
                    count_messages() >= COMPRESS_THRESHOLD?
                           │
                    compress_chat()  ←  active model summarizes oldest N

Your brain compresses 20 years → 2-minute story. 20centAI compresses 1000 messages → 300 tokens.

Bio-inspired token compression.

MIT — fork it, ship it, make it yours.

Built with the help of claude.ai in cooperation with perplexity.ai and deepseek.com.