PolyMind

PolyMind is a self-hosted personal web app that sends a single prompt to multiple LLMs simultaneously and displays every response side by side in real time. It includes a Debate mode where models critique each other's answers across multiple rounds and try to converge, and a Thread Composer that lets you cherry-pick specific outputs to chain into a follow-up prompt. It's a Progressive Web App (PWA) — installable on Android, ChromeOS, and macOS like a native app.

---

Screenshots

Screenshots coming soon — PRs welcome!

Planned: Prompt tab (multi-model grid), Debate tab (rounds), Thread Composer panel, mobile PWA view.

---

Features

• Multi-model prompting — send one prompt to multiple LLMs at once, see all responses side by side
• Real-time streaming — tokens stream into each card as they arrive, per model
• Free + paid models — OpenRouter (most models) and Google Gemini API (direct billing)
• Image input — paste or drag-drop images; vision-capable models receive the image, others get text only with a visual warning
• Session history — scroll through past prompts in the sidebar without page reload
• Copy as Markdown — copy any individual response with one click
• Cost tracking — token counts and USD cost per model, per session, with a resettable counter
• Live OpenRouter balance — shown in the header, updated after every prompt
• Debate mode — 2 or 3 rounds: independent answers → critique & refine → convergence
• Thread Composer — after a debate, cherry-pick responses from any round/model into a new follow-up prompt
• Product Owner role — assign one model as a reviewer/director instead of a participant; useful for coding tasks
• PWA — installable on Android (Chrome), ChromeOS, macOS (Safari)
• Dark theme — hot pink/magenta accent palette, optimized for readability
• Fully self-hosted — nothing leaves your server except calls to the LLM APIs you configure

---

Supported Models (defaults)

Model	Provider	Tier	Input / Output (per 1M tok)	~Prompts per $10
Llama 3.3 70B	OpenRouter	Free	$0 / $0	unlimited*
Gemini 2.0 Flash	OpenRouter	Free	$0 / $0	unlimited*
DeepSeek R1	OpenRouter	Free	$0 / $0	unlimited*
Qwen 2.5 72B	OpenRouter	Free	$0 / $0	unlimited*
GPT-4o mini	OpenRouter	Paid	$0.15 / $0.60	~37,000
Gemini Flash Pro	Google API	Paid†	$0.075 / $0.30	~125,000
Claude Haiku 4.5	OpenRouter	Paid	$1.00 / $5.00	~4,200
Grok 4.3	OpenRouter	Paid	$1.25 / $2.50	~6,700

* Free models have rate limits: ~20 req/min, ~200 req/day per OpenRouter
† Billed to your Google Gemini plan, not OpenRouter credits — cost shown in app is an estimate

Tip: Models are defined in litellm_config.yaml. Add or remove any model supported by LiteLLM without changing application code.

---

Requirements

• Docker and Docker Compose (v2+)
• An OpenRouter account and API key — the free tier works for all free models
• A Google Gemini API key — only needed if you want to use Gemini Flash Pro
• A server or machine running Linux (or any Docker-capable host)
• ~512 MB RAM available for the containers

---

Setup

# 1. Clone the repo
git clone https://github.com/your-username/polymind
cd polymind

# 2. Copy the environment template
cp .env.example .env

# 3. Edit .env and fill in your keys
#    OPENROUTER_API_KEY — from openrouter.ai/keys
#    GOOGLE_API_KEY     — from aistudio.google.com (optional)
#    LITELLM_MASTER_KEY — any random string (internal auth)
nano .env

# 4. Start PolyMind
docker compose up -d

# 5. Open in your browser
open http://localhost:7860

First startup pulls Docker images — allow 1–2 minutes.

---

Accessing from other devices

Replace localhost with your server's LAN IP:

http://192.168.1.60:7860

If you use Tailscale, use your Tailscale IP from anywhere:

http://100.90.190.127:7860

---

Installing as a PWA

Android (Chrome): Open the app URL → tap the ⋮ menu → "Add to Home Screen"

ChromeOS: Same as Android, or look for the install icon in the address bar

macOS (Safari): Open the URL → click Share → "Add to Dock"

Once installed, PolyMind behaves like a native app with its own icon and standalone window. The shell is cached by the service worker, so it loads instantly even on slow connections — but LLM API calls always require internet.

---

Adding or removing models

Edit litellm_config.yaml and add or remove a model block following the existing format:

- model_name: my-new-model
  litellm_params:
    model: openrouter/provider/model-name
    api_key: os.environ/OPENROUTER_API_KEY
    api_base: https://openrouter.ai/api/v1

Then add the corresponding entry to models_config.py's _MODEL_METADATA dict (display name, cost, vision support, etc.) and restart:

docker compose restart

Model strings must follow LiteLLM's format — the provider prefix (e.g., openrouter/) is required.

---

The Debate feature

The Debate tab runs a structured multi-round conversation between your selected models:

• Round 1 — Independent answers: each model responds without knowing what the others said
• Round 2 — Critique & refine: each model sees all Round 1 responses and refines its answer
• Round 3 — Convergence: each model synthesizes everything into a final answer

Because context grows with each round (Round 2 includes all Round 1 responses, Round 3 includes both), a 3-round debate with 4 models costs roughly 5–6× a single prompt. The cost estimate banner shows you the expected cost before you start.

Thread Composer

After a debate, click Build Thread → to open the Thread Composer. It lets you:

1. Check individual responses by round and model to include in a composed context
2. See a live preview of the context as it will be sent
3. Add a follow-up question or instruction
4. Choose which models receive the follow-up (can be different from the debate participants)
5. Click Send to Prompt tab → — the follow-up fires automatically with the composed context pre-loaded

Product Owner role

Right-click any model chip in the Thread Composer to assign it the Product Owner role (crown 👑). Instead of answering the follow-up directly, the PO model reviews all other models' outputs, identifies the strongest approach, flags risks or mistakes, and delivers a clear recommendation. Only one PO is allowed at a time — assigning a new one demotes the previous. This is especially useful for coding tasks where you want one model to review and direct rather than just solve.

---

Checking your OpenRouter balance

Your current OpenRouter credit balance is shown live in the top-right corner of the app. It's updated on page load and after every completed prompt. You can also check it at openrouter.ai/account.

Gemini usage is billed separately by Google and the cost shown in the app is an estimate based on published per-token rates — check Google AI Studio for actual billing.

---

Contributing

PRs and issues are welcome. This is a personal tool built for fun and productivity — if you add something useful, please share it back.

A few guidelines:

• One feature or fix per PR
• Keep dependencies minimal — the goal is a lightweight, self-hostable tool
• Test with at least one free model before submitting

---

License

MIT