Cumbersome vs LLM Convergence

Use Apple Private Cloud Compute, your API keys, or an OpenAI-compatible endpoint. I built Cumbersome because consumer AI apps kept dumbing things down. Cumbersome is the manual transmission of AI apps: more complex, but with far greater control. WHY I BUILT THIS Consumer AI apps hide the controls I want and lock me into one provider. I wanted direct API access, the freedom to pick and compare models, and to pay providers for what I use. This is the app I use every day, so I pay attention and try to get it right. SUPPORTED PROVIDERS - Apple Private Cloud Compute - ChatGPT / OpenAI - Claude / Anthropic - Gemini / Google AI Studio - OpenRouter multi-provider - Vercel AI Gateway multi-provider - Z.ai - OpenAI-compatible custom endpoints, including Ollama, Azure, LM Studio, Groq, Together, and self-hosted backends WHAT YOU GET - A versatile API client. One app, every major provider and model. - Total control. Pick your model, tune the parameters, switch providers mid-conversation. - Compare models head to head. Stream multiple chats in parallel and let an AI judge pick the best. - Bring your own tools. Local browser use, reasoning, web search, image generation. - Rich, editable conversations. Markdown, LaTeX, code blocks, attachments, edit and export anything. - Private by design. iCloud sync across devices. Nothing saved to our servers. - Native and accessible on iPhone, iPad, and Mac. MAJOR FEATURES ADDED IN 2026 New in 2.0.0: Apple Private Cloud Compute. No API key. Requires iOS/iPadOS/macOS 27, an Apple Intelligence device, and internet. Apple usage limits apply. 2026-09-21: Independent iPad chat windows. 2026-09-14: Read replies aloud as they arrive. Save and Restart after editing. 2026-08-29: Select multiple chats. Mac export/import of all chats. 2026-08-21: Background replies on iOS 26. Question Cards for clarification. 2026-08-18: Web Agent: a shared, interactive browser. 2026-08-06: Highlight passages and attach notes for your next reply. 2026-08-03: Local Python analysis, charts, Mermaid, Markdown export. 2026-07-31: Gemini video. Responses API for custom providers. 2026-07-28: Parallel browser and MCP tools. 2026-07-23: Mind/Meld combines three AI responses into one. 2026-07-18: Full Simplified Chinese support. 2026-07-12: Ghost/Write drafts your next turn. 2026-07-09: Remote MCP servers and tools. 2026-07-02: Chat themes and reading fonts (Serif, Rounded, Mono). 2026-06-28: Folder system message presets. 2026-06-24: Prompt caching, with usage across providers. 2026-06-17: Reusable system presets per conversation. 2026-06-16: Offload long system messages to iCloud Drive. 2026-06-15: Toggle custom API parameters. 2026-06-11: Compact chats to lower API costs. 2026-06-07: Run JavaScript and HTML. 2026-06-04: OpenRouter PDF parsing and search engine options. 2026-05-12: Quick Prompt shortcut and App Visibility on Mac. 2026-05-09: Folders to group and organize chats. 2026-05-03: Collapsible sidebar sections, synced across devices. 2026-04-22: Image generation with OpenAI and Gemini. 2026-04-18: Local browser search and browsing. 2026-04-11: Headless browser mode. 2026-04-03: Searchable model picker. 2026-03-15: Extended thinking and reasoning across providers. 2026-03-09: Auto-delete chats after a set number of days. 2026-02-28: Custom OpenAI-compatible providers. 2026-02-25: Only Search When Requested. 2026-02-24: Image paste, attachment previews, PDF thumbnails. 2026-02-21: Siri and Shortcuts. 2026-02-20: Vercel AI Gateway and model filters. 2026-02-06: Voice Control, Larger Text, higher contrast. 2026-01-11: VoiceOver, Dynamic Type, keyboard navigation. 2026-01-06: OpenRouter integration. THE DEAL Core features are free. An optional Cumbersome Pro subscription adds advanced Project tools and supports ongoing development. You pay API providers directly for usage. I'm Peter, an indie developer in Reno, Nevada, out among the high desert tumbleweed. No VC money. No growth hacking. Just a tool I wanted to exist.

One AI model can be confidently wrong. LLM Convergence asks several at once and makes them reconcile. Pose a question and LLM Convergence sends it, in parallel, to a panel of leading AI models. A model you pick acts as the leader: it breaks every answer down into individual claims and reconciles them over a few rounds. What comes back is a single answer the panel converged on - plus a convergence score, and the specific points they disagreed on surfaced instead of buried. HOW IT WORKS - Ask in parallel - your prompt goes to several models at once, across OpenAI, Anthropic, Google, and fast open-model hosts. - Reconcile - a leader model decomposes each answer into claims and reconciles them round by round, so agreement is earned, not assumed. - See convergence - read one reconciled answer with a convergence score, and inspect the claims that didn't line up. TWO CONSENSUS MODES - Single judge - a leader model judges and reconciles each round. - Claim-commit voting - claims commit by panel quorum and stay committed (inspired by RAFT). GROUND ANSWERS IN THE WEB (OPTIONAL) Turn on grounding and LLM Convergence runs a web search before the panel answers, then puts the cited results in front of every model - so they reason over real sources instead of memory alone. Grounded answers come back with tappable [S1]-style citations and a Sources list, a grounding score next to the convergence score, and an honest "couldn't ground this" when a search turns up nothing. It uses a frontier key you already have (Claude, Gemini, or an OpenRouter model) - no separate search key, and it works on every platform. FACT-CHECK MODE Ask LLM Convergence to fact check something and it switches from answering to verifying. Paste a link and it opens the page - or check pasted text, a bare claim, or the answer it just gave you. It pulls out the checkable claims, runs a separate web search for each one, has the whole panel weigh the evidence, and returns a verdict per claim - supported, refuted, partly supported, or unverifiable - with citations you can tap. It reads how you phrased the question, so there is no mode to switch on. Fact-check mode needs a model that can search the web. BRING YOUR OWN KEYS Use your own provider API keys. They're stored securely in your device's keychain and used only to call the providers you choose - never sent to us. One key per provider: add more models from a vendor you've already set up without re-pasting its key, or override the key for a specific model. Mix and match models, run more than one model from the same vendor, and pin whichever one leads. PRIVATE BY DESIGN LLM Convergence is a pure client - there's no account and no server of ours. - No analytics, no telemetry, no ad SDKs. - Prompts go straight from your device to the providers you configured. - Chat history is saved locally and stays on your device. Delete any conversation anytime. YOU'RE IN CONTROL Choose which models sit on the panel, which one leads, how many rounds run, and the convergence threshold. Multi-turn chat with local history and Markdown-rendered answers. Run more than one conversation at a time - a chat keeps working in the background while you read another. Supported providers: OpenAI, Anthropic (Claude), Google (Gemini), and OpenAI-compatible hosts including Groq, Fireworks, Cerebras, and OpenRouter - the last reaching many models behind a single key - with free open models available. Note: LLM Convergence requires your own API keys from the providers you want to use. Provider usage is billed by those providers per their own pricing.