Polychat AI Platform

Polychat is an experimental multi-model AI assistant platform. It brings chat, agents, retrieval, generated media, code assistance, and workflow-style apps into one Cloudflare-backed API with React web and iOS clients (In Development).

Check out my write up on this project here. I've also launched a version of this to try out at polychat.app.

Note

Please note that this project is still in active development so there are a few features that are not yet fully working or fully imagined. I'm also constantly testing stuff and re working, this is meant to be a playground for me.

Quick Links

• API Documentation - Live OpenAPI schema at api.polychat.app/openapi (served from apps/api/src/index.ts, with tags in apps/api/src/openapi/documentation.ts)
• Example Chats - See what's possible
• Getting Started - Set up your own instance
• Features - What's included

Example Chats

Here are some example chats that you can try out:

• Generation of a new React component
• Code generation and execution
• Web search integration
• Perplexity Deep Research
• Combined artifact previews
• Markdown formatting
• Search Grounding
• Multi-model responses
• Saved memories (RAG)
• Retrieved memories (RAG)
• Multi Step Tool Calls
• Multi Step MCP Calls
• Agent to agent delegation
• Image Generation
• Council

What's Included

This monorepo contains:

• API - OpenAI-style API for chat, agents, models, tools, memories, media, realtime sessions, payments, and webhooks
• Web App - React-based PWA frontend for chat, shared conversations, generated artefacts, and focused app workflows
• Sandbox Worker - Automated coding worker built on Cloudflare Sandboxes
• Metrics Dashboard - Usage analytics and monitoring
• Training Worker - Internal Worker for API-managed model training and deployment jobs
• Mobile App - iOS application (In Development) (TestFlight)

Features

Core API Features

• Chat Completions - OpenAI-style chat with streaming, tools, multimodal message parts, stored conversations, sharing, feedback, and token counting (tag: chat)
• Multi-provider Model Library - Anthropic, OpenAI, Google, Mistral, Meta, Bedrock, Groq, DeepSeek, Perplexity, Replicate, xAI, and many more through the provider registry (tag: models)
• Code Assistance - FIM completions, next-edit suggestions, apply-edit support, and sandboxed implementation workflows (tag: code)
• Agents & Delegation - Custom agents, shared agents, MCP server integrations, agent-to-agent delegation, and approval-aware tool flows (tag: agents)
• RAG & Memories - Vector-based context retrieval, memory groups, synthesis tasks, and namespace-aware retrieval with Cloudflare Vectorize (tag: memories)
• Tools & Dynamic Apps - Searchable function tools, dynamic app execution, reusable app responses, and app sharing (tags: tools, dynamic-apps, apps)
• Media & Realtime - Uploads, audio transcription and speech, image/video/music generation, screenshots, OCR, and realtime transcription sessions (tags: uploads, audio, realtime)
• Guardrails & Auth - Content safety checks, OAuth, magic links, passkeys, JWT, API keys, plan checks, Stripe subscriptions, rate limiting, CSRF, and security headers (tags: guardrails, auth, plans, stripe)

Web App Highlights

• Automated Model Routing - Smart model selection
• AI Council - Multi-perspective debate mode with selectable council members and synthesis
• Canvas - Multi-model image and video generation with shared prompt controls
• Replicate App - Run media and multimodal predictions from Replicate models
• Sandbox App - Connect GitHub repositories and stream implementation, review, test, or bug-fix runs
• Strudel Music Patterns - Generate and edit code-based music patterns
• Podcast, Article, and Note Workflows - Transcribe, summarise, report, format, and regenerate content
• Drawing Apps - Create drawings and ask AI to enhance or guess them
• Web LLM Support - Local/offline chat mode in the web app
• Metrics Dashboard - Usage, model performance, costs, and user activity monitoring

Automated Coding with Sandbox Worker

The Sandbox Worker uses Cloudflare Sandboxes to run AI-powered code generation tasks against GitHub repositories. This enables automated feature implementation triggered via GitHub comments or the web UI.

I wrote more about this in my blog post here.

How it works:

1. Install the GitHub App on your repository
2. Start an implementation, review, test, or bug-fix run from the web UI or GitHub workflow
3. The AI generates a plan and executes commands in an isolated sandbox
4. Progress streams back to the app while results are prepared for review

Key capabilities:

• Isolated command execution with security restrictions
• Real-time progress streaming via SSE
• Git operations (clone, branch, commit)
• Web UI for installation, run history, instructions, pausing, cancellation, and task tracking

See all features in the OpenAPI reference → api.polychat.app/openapi

Usage Limits

Polychat is configured with usage limits to prevent abuse. These limits are as follows:

• 10 standard messages per day for unauthenticated users
• 50 standard messages per day for authenticated users
• 200 pro tokens per day for authenticated users

Pro tokens are calculated based on a multiplier of the cost of the model. For example, if a model costs $0.01 per 1000 input tokens and $0.05 per 1000 output tokens, then the pro token limit is 200 * (0.01 + 0.05) / 2 = 6.

This equates to around:

• Expensive models (9x): ~22 messages
• Mid-tier models (3x): ~66 messages
• Cheaper models (1-2x): 100-200 messages

If you are providing your own service and would like to change these limits, you can do so by changing the USAGE_CONFIG object in the apps/api/src/constants/app.ts file.

Setup and Installation

Getting Started

1. Clone the repository

2. Install dependencies

   pnpm install

3. Configure environment variables:

   • Copy .dev.vars.example to .dev.vars in all the apps directories that have them.
   • Copy wrangler.jsonc.example to wrangler.jsonc in all the apps directories that have them.
   • Adjust with your API keys and configuration values.

4. Start the development servers:

   # Start the API and web app
   pnpm run dev
   
   # Or start individual apps
   pnpm run dev:app
   pnpm run dev:api
   pnpm run dev:metrics

Deployment

The applications are designed to be deployed to Cloudflare:

# Deploy all applications
pnpm run deploy

# Deploy individual applications
pnpm run deploy:app
pnpm run deploy:api
pnpm run deploy:metrics

API Documentation

The complete API documentation is generated by the API worker and served at api.polychat.app. Route registration lives in apps/api/src/index.ts, with tag descriptions in apps/api/src/openapi/documentation.ts.

Quick Example

curl https://api.polychat.app/chat/completions \
  -H "Authorization: Bearer YOUR_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "claude-3-5-sonnet-20241022",
    "messages": [
      {"role": "user", "content": "Hello!"}
    ]
  }'

View full API reference → api.polychat.app

Database Management

The application uses a Cloudflare D1 database with Drizzle ORM for schema management and migrations.

Running Migrations

cd apps/api
# Migrate to the local database
pnpm run db:migrate:local

# Migrate to the preview database
pnpm run db:migrate:preview

# Migrate to the production database
pnpm run db:migrate:prod

To generate a new migration, run:

pnpm run db:generate

Monitoring and Analytics

The metrics application provides dashboards for monitoring:

• API usage and performance
• Model performance and costs
• User activity and engagement

Access the metrics dashboard at metrics.polychat.app.

Contributing

Contributions are welcome! Please feel free to submit a Pull Request.

License

This project is licensed under the terms of the license included in the repository.