Ramelli
Thirty models in one conversation.
Ramelli is an AI chat workspace. Start with a quick question, switch models without losing the thread, or hand it to Deep Research and watch it plan, search and check its own answer.
- Models
- 30+ from 9 providers
- Modes
- Standard, Web search, Pro Search, Deep Research
- Built with
- Next.js 14, TypeScript, Convex, Redis, Turborepo, OpenRouter, Clerk, Vercel AI SDK, Langfuse, Sentry, PostHog, OpenTelemetry, Cloudflare R2, Zustand, Radix UI
The idea
Ramelli puts thirty-plus models from nine providers behind one conversation, routed through a single OpenRouter layer. You can switch models mid-thread and keep the context.
It’s local-first: your data stays on the device, with optional real-time sync through Convex. All four modes, from a quick answer to multi-step Deep Research, share the same streaming interface with the model’s reasoning visible.
The research engine
Each request runs as a graph of tasks in a custom, type-safe workflow engine, @repo/orchestrator, with dependencies, retries, timeouts and parallel fan-out.
A router picks one of four flows. Deep Research plans its searches, runs them in parallel, and loops a reflector back to the planner until the answer holds up. Pick a mode below to follow a request through it.
- 01router
Entry point: routes each message to the right chat mode.
- 02refine-query
Analyses and clarifies the research question.
- 03planner
Plans strategy and generates search queries.
- 04web-search
Runs searches in parallel and processes content.
- 05reflector
Evaluates progress; loops back if gaps remain.
- 06analysis
Synthesises findings before the report is written.
- 07writer
Writes the comprehensive, cited report.
- 08suggestions
Proposes follow-up questions to keep exploring.
On every screen
The models
Nine providers, reached through one API.
- OpenAI
- Anthropic
- DeepSeek
- xAI
- Alibaba
- Moonshot
- Zhipu
- MiniMax