Roundtable AI · macOS · Free
Where AI Minds
Meet.
Stop asking one AI. Ask a room. Seat up to 8 agents — each with a different persona and model — and watch them disagree, debate, and converge on your question.
macOS 13 Ventura or later · Apple Silicon · Local-first, no cloud required
One AI gives you
an answer. Eight give you the truth.
Single-model AI assistants tend to tell you what you want to hear. They're optimised for helpfulness, which often means optimised for agreement. Put eight models at a table and that dynamic breaks.
One voice. One perspective. One blind spot.
- Single model gravitates toward the most statistically likely answer
- Optimised for helpfulness, which often means optimised for agreement
- No genuine disagreement — no one to challenge the premise
- You don't know what it missed
Eight voices. Real disagreement. Conclusions you can trust.
- Each agent reads what the others said before responding
- Devil's Advocate finds the flaw. Analyst cites the numbers. Pragmatist asks who does it.
- Structured synthesis at the end: what they agreed on, where they didn't
- You stay in control — intervene, redirect, challenge any agent at any point
One conversation.
Multiple minds.
A structured debate engine between multiple models — giving you richer analysis, genuine disagreement, and synthesised conclusions.
Multi-Agent Orchestration
A round table, not a chatbot
Most AI tools give you one model, one voice. Roundtable AI seats up to 8 agents around a shared conversation — each with a distinct persona, model, and perspective. They respond to each other, challenge assumptions, and build on what came before.
- Seat 2 to 8 agents per session — mix and match models freely
- Round Robin, Moderator, Vote, and Dynamic turn modes
- 5 discussion phases from Opening through Synthesis
- Real-time token streaming — watch each agent think
- Full session history saved automatically, resume any time
Director Controls
You're always in the room
Roundtable AI doesn't run on autopilot. Pause after any turn, inject context as the Director, challenge an agent, redirect the conversation mid-session. When you're satisfied, trigger synthesis and get a structured conclusion.
- Pause and resume sessions at any point without losing state
- Inject "Director" messages visible to all agents as shared context
- Challenge, elaborate, spotlight, constrain, or skip any agent
- Force conclusion: pick which agent synthesises the debate
- Export full sessions as Markdown, JSON, HTML, or PDF
Model Flexibility
Local, cloud, or both — you choose
No vendor lock-in. Run fully local GGUF models on-device for complete privacy. Connect to OpenRouter for GPT-4o, Claude, Mistral, Llama, and 100+ more. Each agent picks its own model — mix freely in the same session.
- Local GGUF — Metal GPU acceleration on Apple Silicon, no internet
- OpenRouter: one API key, 100+ models including GPT-4o and Claude
- Ollama: auto-detect a running local Ollama instance
- Hardware-aware: RAM guidance so you never overload your Mac
- Per-agent model selection — each agent is independent
Four ways to run
a round table
Each mode changes who speaks when and why. Pick the one that fits the problem.
Round Robin
Fixed rotation. Every agent speaks in sequence. Equal participation, predictable structure.
Moderator
A designated moderator agent selects who speaks next based on relevance and discussion flow.
Vote
Agents vote on who should speak next. The majority decides. Surfaces who the group considers most relevant.
Dynamic
Adaptive selection based on the current discussion phase. Shifts behaviour from opening through synthesis automatically.
Sessions that go somewhere
Every session moves through structured phases. Agents receive phase-appropriate prompting, so the conversation naturally shifts from exploring to converging.
Opening
Agents introduce their perspective on the goal.
Exploration
Diverse angles surface. No convergence pressure yet.
Deepening
Specific threads get pressed. Assumptions challenged.
Convergence
Common ground mapped. Persistent disagreements named.
Synthesis
Structured conclusion drawn by the designated agent.
Every intervention you need,
exactly when you need it
Roundtable AI doesn't run on autopilot. Eight controls let you shape the conversation at every step without breaking the flow.
Pause / Resume
Stop the session mid-turn. Resume exactly from where it left off, state fully preserved.
Inject
Add a Director message all agents read as shared context before their next turn.
Elaborate
Ask a specific agent to expand on their last point without moving to the next turn.
Challenge
Force an agent to reconsider their position or defend it against a specific objection.
Spotlight
Override turn strategy and pick exactly who speaks next, regardless of mode.
Skip
Pass the current agent's turn without generating a response. Move the conversation forward.
Constrain
Apply a temporary rule to all future turns (e.g., "respond in bullet points only").
Conclude
Trigger synthesis from the agent of your choice. Get a structured conclusion immediately.
A round table running
in under 60 seconds
Three steps. No configuration files. No Python. Just open the app and go.
Connect your models
Point Roundtable AI to your local GGUF files, enter your OpenRouter API key, or connect a running Ollama instance. The app reads your hardware and recommends how many models you can run at once. One-time setup.
Seat your agents
Choose from built-in personas or create your own. Assign a model to each agent, pick a discussion mode, set a goal. Use a pre-built template to skip setup entirely.
Watch them think
Each agent reads the full conversation before responding. Inject context, challenge assumptions, redirect the thread, or force a conclusion. Export the result when you're done.
Start with a template,
not a blank page
Pre-configured round tables with hand-picked agents, personas, and discussion settings. Pick a template, set your goal, start.
Think
Startup Pitch Review
Analyst, Devil's Advocate, and Finance Expert stress-test your pitch from every angle before investors do.
Think
Red Team / Blue Team
Two opposing teams attack and defend a strategy. Forces you to see both sides before committing.
Think
Hiring Committee
Multiple evaluators review a candidate profile and surface different perspectives on fit and risk.
Explore
Philosophical Inquiry
Diverse thinkers explore ethical dilemmas and foundational questions from multiple schools of thought.
Explore
Policy Debate
Legal, economic, and social perspectives clash over a proposed policy. Ideal for understanding trade-offs.
Create
Writers' Room
Creative Thinker, Marketer, and Pragmatist brainstorm campaign ideas, story angles, or product names.
Voices ready to go
Every persona ships with a crafted role description, debate style, and response preference. Use them as-is or customise everything — name, avatar, system prompt, debate style, temperature.
The Analyst
Data-driven, systematic. Cites evidence before opinions.
Devil's Advocate
Challenges every assumption. Finds the flaw in the plan.
The Optimist
Finds the upside. Energises the group with possibility thinking.
The Pragmatist
Grounds the debate in execution reality. "But who does it?"
Creative Thinker
Lateral, unconventional. Spots angles nobody else considered.
Fact Checker
Questions sources and assumptions. Keeps the debate honest.
Moderator
Structures the debate. Decides who speaks next in Moderator mode.
Summariser
Distils threads into clear takeaways. The ideal conclusion agent.
Legal Expert
Flags compliance issues and liabilities. Essential for decisions with risk.
Engineer
Thinks in systems and trade-offs. Challenges feasibility claims.
Marketer
Obsesses over positioning, messaging, and the customer story.
Finance Expert
Models the numbers. Questions unit economics and runway assumptions.
Custom Persona
Build your own with any name, avatar, role, expertise tags, and system prompt.
Start free. Go further with Pro.
The free tier is genuinely useful. Pro removes the limits and unlocks the advanced capabilities.
$0
No account. No credit card.
- Up to 8 agents per session
- All discussion modes and personas
- OpenRouter and Ollama connections
- 1 local GGUF connection
- Limited sessions per day
Pro
One-time purchase. Yours forever.
- Everything in Free
- Unlimited sessions
- Multiple local GGUF connections
- Web research agent (live internet injection)
- Advanced auto-synthesis and insight extraction
- Cross-session memory and context summaries
Bring your own models
Roundtable AI ships with no bundled models, keeping the install lightweight. Connect what you already have — or what you prefer.
Local GGUF Models
Point to any .gguf file on your Mac. Inference runs entirely on-device with Metal GPU acceleration on Apple Silicon. No internet. Complete privacy.
OpenRouter
One API key unlocks GPT-4o, Claude, Llama, Mistral Large, Gemma 2, and 100+ more. Mix cloud agents with local ones in the same session — each agent is independent.
100+ ModelsOllama
Auto-detect a running Ollama instance and use any of your downloaded models instantly. For users already in the Ollama ecosystem — no extra configuration needed.
AvailableFrequently asked
Get Roundtable AI
Free on the Mac App Store. No account, no subscription, no telemetry. Your conversations stay on your machine.