Prepended to every prompt in sessions belonging to this project.
Files uploaded here are injected into every prompt in this project's sessions. Any file type (10MB cap).
Move "" to:
LLM Council
Connected
Showing all messages
LLM Council
Five-model consensus engine. Claude, ChatGPT, Gemini, Grok, and DeepSeek debate, peer-review, and synthesize the best answer.
Session files:
×
── Council Modes ──
Debate — Council
Consensus — One Answer
Brainstorm — Idea Generation
Research — Deep Dive
Verify — Check Work
Clarify — Ask First
── Executive Modes ──
Assignment — Plan + Execute
Brief — Exec Assistant
Strategize — Strategic Analysis
Review — Life/Biz Org
Coordinate — Daily Rhythm
── Operations Modes ──
Communicate — Draft Message
Concierge — Arrange It
Monitor — Project Watch
── Direct Models ──
Parallel — All Expert
Parallel — All Assistant
Parallel — All Intern
Claude
Opus 4.6 — Expert
Sonnet 4.6 — Assistant
Haiku 3.5 — Intern
ChatGPT
GPT-5.4 — Expert
GPT-5.4 mini — Assistant
GPT-4o — Intern
Gemini
Gemini 3.1 Pro — Expert
Gemini 3 Flash — Assistant
Gemini 2.5 Flash — Intern
Grok (xAI)
Grok 3 — Multi-Agent
Grok 4.2 — Reasoning
Grok 4.1 Fast — Intern
DeepSeek
DeepSeek — Expert
DeepSeek — Assistant
DeepSeek — Intern
General
App-wide preferences. Council member assignments + arbiter + coder live in the Council pane (admin only).
Which mode to use when you first open the app.
Prefer responses in a specific language (leave blank for English).
Adjust font and icon sizes across the entire interface.
AA100%
Personalization
Customize how the Council responds to you.
The more you put here, the more the council gets to know you. All fields optional. Auto-injected into every prompt — no more "what's your name?" intros.
Prepended to every prompt — set tone, formatting, behavior preferences. Stored server-side so it works on every device.
Persistent notes about you that the council reads on every prompt — across all sessions. Council members can also append to this when they learn durable facts about you. Edit freely.
v2 = modern minimalist (default). v1 = original baseline. Switch is instant — no reload needed.
Stacked = traditional vertical. Columns = council member responses side-by-side, truncated with "Show full" toggle (handy when you want to compare answers).
Get notified when a council round completes or the arbiter finishes — even when this tab is closed (works on iOS Safari 16.4+ as a Home Screen app).
LLM Models
Choose which models to use for each tier. Use Refresh to pull latest models from each provider.
Claude / AnthropicCLI
ChatGPT / OpenAICLI
Gemini / GoogleCLI
Grok / xAIAPI
DeepSeekCLI
QQwen / AlibabaAPI
MMiniMaxAPI
Saves settings AND tells every connected user to reload their dropdown labels immediately. Settings are global — every user sees the same model assignments.
Data & Files
Manage global context files injected into every prompt.
These files are injected as context into ALL prompts for ALL sessions.
Account
Manage your account and security.
My Connections
Connect your own accounts so the council can act on your behalf — read your Drive, draft your Gmail, check your Calendar. Each connection is private to you and your tokens are encrypted at rest. Disconnect any time.
Loading…
Task Blueprints
Review and edit a task blueprint before approving it for execution. Each blueprint is private to you. Expand one to edit its objective, summary, constraints and deliverables, then Approve when it's ready to run.
Loading…
Audit Log
A timeline of every external action taken on your behalf — what was done, who/which agent did it, the tools and sources used, and its approval state. Newest first.
Showing the most recent events.
Loading…
Role Debate
When you run a Role Debate, each council member argues from an assigned role. Pick the role you want each model to play, or leave it on Auto to use the research-backed best fit for that model. Your choices are saved to your account and used as the default for every Role Debate you run.
Loading…
Applies to your future Role Debates.
Brainstorm
When you run Brainstorm, the council generates ideas across a divergent round, then builds on each other collaboratively, then a Brainstorming Lead synthesizes + ranks them. Choose how members ideate by default — Parallel (each member ideates independently) or Role-based (each member gets an ideation hat: Visionary, Pragmatist, Customer Voice, Constructive Contrarian, Domain Expert, Synthesizer, Wildcard). You can still flip this per-brainstorm from the composer.
Saved instantly. This is your default for every brainstorm; the composer toggle overrides it per-run.
Task Routing
Assign the best model for each task type. The system auto-detects task type from prompts and routes to the assigned model when using individual queries.
Code Generationwriting new code
Code Reviewbugs, security, quality
Planning & Architecturesystem design, roadmaps
Creative Writingstories, copy, marketing
Data AnalysisCSV, charts, SQL
Summarizationcondense documents
Math & Reasoninglogic, proofs, calculations
Researchfact-checking, deep analysis
Vision / Image Analysisunderstand uploaded images
General Chatconversation, Q&A
LLM Connections
Configure API keys and connection modes. Auto = API if key exists, else CLI.
ChatGPT / OpenAI
Ensure CLI is logged in. Need to re-auth?
Gemini / Google
Ensure CLI is logged in. Need to re-auth?
Claude / Anthropic
Ensure CLI is logged in. Need to re-auth?
Grok / xAI
Grok always uses API. Key is pre-configured.
DeepSeek
CLI mode uses opencode for direct file system access. DeepSeek participates in council debates only when a key is set.
QQwen / Alibaba
DashScope International (Singapore) endpoint. Qwen joins council debates only when a key is set.
MMiniMax
MiniMax International (minimax.io). MiniMax joins council debates only when a key is set.
Council
Every dropdown below is populated live from each provider's API. Click Refresh All to repull the latest model lists. The "fallback" runs when a council member errors out, so the debate doesn't degrade.
Claude seat
ChatGPT seat
Gemini seat
Grok seat
DeepSeek seatkey required
Qwen seatkey required
MiniMax seatkey required
Synthesizes the final answer. Pick from any provider's full model catalog.
After Arbiter synthesizes, this model implements code with direct file access (CLI mode). Set "None" to skip.
When a council member errors / times out / returns empty, that seat is retried with this model. Pick something cheap and reliable.
When V2 (efficient) pipeline is active, this model condenses the prompt + history into the brief that the council reads.
The "Brainstorming Lead" that ranks + categorizes everyone's ideas into the final list. Leave on default to use GPT-5.5.
MCP Servers
Register remote MCP servers. Each server can be allowlisted per council member. Wired into Claude (via anthropic-beta: mcp-client-2025-11-20), OpenAI Responses API, and Grok Agent Tools. Examples: filesystem, github, gmail, slack, playwright, custom.
⚠ Safety: MCP servers can take real-world actions (send email, run code, modify files). Only register servers you trust. Auth tokens are stored server-side and never exposed in the UI.
User Management
Create and manage accounts. Three roles available: Admin (full access), Developer (chat + code modes), User (chat only).
LLM Council — Guide
What this is
LLM Council orchestrates five AI providers — Claude, ChatGPT, Gemini, Grok, and DeepSeek — to debate, peer-review, and synthesize the strongest answer to your question. You pick a question, the council goes through a structured back-and-forth, and you get one consolidated answer (with the option to peek at any individual member's response).
It's much more than debate. The council runs in any of 13+ modes — research, verification, brainstorming, an executive briefing, drafting a message, arranging plans, monitoring your open work, and more — and can act on your connected accounts (Drive, Gmail, Calendar) when you allow it. See the Modes and Connections tabs.
The composer (＋ menu and ↑ send)
The chat bar is a single iOS-style composer — the same on web and the iPhone app:
＋ button (leading edge) — opens a menu with Mode (pick which mode/model to run), Export session, Attach (files or QR scan), and Image (AI image generation).
↑ circular send arrow (trailing edge) — sends your message. During a live debate it becomes Interject; in a single-model chat it just sends.
Mode-aware placeholder — the input's prompt text changes to reflect the active mode (e.g. "Convene the council…" vs "Ask Opus 4.6…"), so you always know what your next message will do.
There's no longer a left tool-rail — the ＋ menu owns those actions at every screen size.
How a typical debate works
Leave the mode on Debate — Council (the default), or pick another via the ＋ → Mode menu
Type your question and tap the ↑ send arrow
The council runs through stages — you see each member's response stream in
After the automatic rounds, you decide: ↻ New Round for more refinement, or ️ Send to Arbiter for the final synthesis
Once the Arbiter speaks, you can keep adding rounds or chat about the answer in Discussion Mode
Operating modes
Open the ＋ menu → Mode (or the chat-bar selector) to choose how the next message is handled. Modes are grouped: Council, Executive, Operations, and Direct (single/parallel models). Pick the one that fits the job.
Council modes
DebateAll five members debate, peer-review, and synthesize one consolidated answer. The default.
BrainstormDivergent idea generation across the council, then synthesis and ranking of the best ideas.
ResearchDeep research — separates known / plausible / speculative, with citations.
VerifyVerifier gate — checks a work product (text, plan, or code) for correctness and consistency against your request.
ClarifyClarifying engine — surfaces the questions worth answering before acting, instead of guessing.
Executive modes
AssignmentGoal-driven: a planner decomposes your goal into a task blueprint, you approve it, then the council executes. Review/approve in Settings → Task Blueprints.
BriefExecutive-assistant mode — returns a decision-ready briefing on the request (chief-of-staff layer).
StrategizeStrategic analysis — situation, opportunities, risks, and recommended moves (CEO-mode deep reasoning).
ReviewLife & Business organization — reviews and organizes across your domains.
CoordinateOperational coordination — your daily rhythm across all open work.
Operations modes
CommunicateDrafts a message in your voice for a channel (email, etc.). Can send through your connected Gmail when you approve.
ConciergeResearches and arranges options against your preferences and budget — then lays out the choices.
MonitorProject monitoring — scans your open work for stalls and things needing attention.
Direct (no council)
Parallel — All Expert / Assistant / InternFan a single prompt to all five providers at the chosen tier — every member replies, no synthesis.
Single model15 entries (3 tiers × 5 providers) — pick one for a one-on-one (e.g. Opus 4.6, GPT-5.4, Gemini 3.1 Pro).
Executive and Operations modes that take real actions on your behalf run through a task blueprint you approve first, and every external action is recorded in your Audit Log — see the Tasks & Audit tab.
Rounds — 2 automatic, then your call
The V2 pipeline runs 2 rounds automatically with no user intervention (you can still interject mid-debate to steer/augment). After Round 2 completes, the decision card appears: ️ Send to Arbiter, or ↻ Add Another Round. There's no cap on rounds — keep clicking ↻ New Round as many times as you want until you're satisfied.
Decision Card
After Round 2 a card appears with two buttons: Run Arbiter or Add Another Round. The card stays put if you close the tab and come back.
Persistent buttons in the bar
The bar at the top of the chat also has ↻ New Round and ️ Send to Arbiter for quick access.
Round refinement prompt (optional)
Type something in the input box BEFORE clicking ↻ New Round and that text becomes your refinement request for the next round. Empty = generic refinement.
Pick up where you left off
Refresh, close the tab, come back tomorrow — your decision card and Round +1 button are exactly where you left them.
Interjecting mid-debate
You can drop a message into the input ANY time during a debate — while members are streaming OR while you're at the round-vs-arbiter decision point. The Send button changes to Interject (with megaphone icon) and the placeholder updates to confirm.
The interjection is queued — it does NOT start a new debate.
The council picks it up at the next stage and either steers the discussion to address it or adds it as additional context (their judgment).
You can interject multiple times.
Your interjection is saved to chat history (with the [INTERJECTION] marker) so you can scroll back and see what you said.
Discussion Mode
Activates automatically after a debate completes. The bar at the top of the chat shows filter buttons (All / Claude / ChatGPT / Gemini / Grok / DeepSeek) plus the round controls.
Filter the transcript
Click Claude (or any individual) and only that member's messages stay visible — your prompts and system messages always stay. Click All to unfilter.
Talking to a single member
Pick that member's specific model from the chat-bar dropdown (e.g. "Claude Opus 4.6") — your next prompt goes only to them. The filter buttons are view-only; they don't change who you're talking to.
@mentions
Type @claude, @chatgpt, @gemini, @grok, or @deepseek at the start of a prompt to address that member directly.
Chat-bar model selector
Open it from the composer's ＋ → Mode menu (it's also the dropdown next to the input). Pick what you want the council to do for the next prompt. Filter-as-you-type input at the top — narrow down 200+ models fast. For the full list of council/executive/operations modes, see the Modes tab.
Top-level options
Council Debate — All (default) → V2 pipeline with all 5 members
Parallel — All Expert → all 5 Expert-tier models reply, no synthesis
Parallel — All Assistant → Assistant-tier of every provider
Parallel — All Intern → Intern tier of every provider (cheap fan-out)
Individual models → 15 entries (3 tiers × 5 providers), pick one for a one-on-one
DeepSeek(joins council debates only when key is set)
…
Use the chat-bar dropdown's filter input to find any specific model fast.
My Connections (per-user)
Connect your own accounts so the council can act on your behalf — read your Google Drive, draft your Gmail, check your Calendar. Open Settings → My Connections and connect each service via Google OAuth.
Per-user, private — your connections are yours alone. Your tokens are encrypted at rest; another user's council never sees them.
Google Drive — the council can read/search your files for context.
Gmail — read for context and draft (or, with your approval, send) messages in Communicate mode.
Google Calendar — read your schedule for coordination and concierge tasks.
Disconnect any time — revoking a connection stops all access immediately.
Syncs across web + app
Connections are stored against your account on the server, so they're available the same whether you're on the web app or the native iOS / TestFlight app — connect once, use everywhere.
Actions taken through a connection always run through a task blueprint you approve, and are recorded in your Audit Log — see the Tasks & Audit tab.
Task Blueprints
When you use a mode that does something (Assignment, Communicate, Concierge, Coordinate, etc.), the council first drafts a task blueprint — a plan of what it intends to do. You review it before anything runs.
Open Settings → Task Blueprints. Each blueprint is private to you. Expand one to edit its objective, summary, constraints, and deliverables, then Approve when it's ready to execute. Nothing acts on your accounts until you approve.
Audit Log
Open Settings → Audit Log for a timeline of every external action taken on your behalf — newest first. Each entry shows:
What was done and which agent/member did it
The tools and sources used (e.g. which connection)
Its approval state
Hit Refresh to pull the most recent events. The blueprint-then-approve flow plus the audit timeline mean you always stay in control of what the council does for you.
Projects (workspaces)
Group related sessions under a folder with shared instructions and files. Click ＋ New in the sidebar Projects section to create one.
Per-project settings (click on any project)
Name — rename freely
Custom instructions — prepended to every prompt in this project's sessions
Project files — uploaded files (any type, 10MB cap each) injected into every prompt
"Also use global instructions / memory / files" toggle — when OFF, the project ignores global context entirely (only project files + project instructions apply). When ON, both layered together. User memory always applies.
Delete — sessions become top-level (NOT deleted), project files are wiped
Moving sessions
Hover any session row → click the folder-arrow icon → pick a project (or "Top-level" to remove from any project). The session moves under the chosen project's expandable folder in the sidebar.
Hierarchy
Click a project name to expand/collapse its sessions. Top-level sessions (no project) live in the Sessions section below.
Web access (all 5 members)
Every council member can read the live web. Two ways to trigger it:
1. Paste a URL
Drop any https://... link into your prompt. Up to 5 URLs per message. Example:
Check this page and tell me what you think:
https://example.com/article
2. Ask them to search
Use phrases like "search the web for X", "google X", "look up X", "find online X", or "web search for X". Example:
Search the web for the latest Pokémon TCG set release dates
Global context files + custom instructions — Settings → Data & Files; Personalization → Custom Global Instructions
Session files — uploaded per-session via the chip bar above the input
Inline attach
button or drag-drop into the chat area to attach files to your message.
User memory (per-user, persistent)
Editable at Settings → Personalization → User Memory. Used as context on every prompt across all sessions. The council may add durable facts about you over time.
Canvas panel
Header → Canvas button. Code blocks from AI responses are auto-extracted. Save, Copy, or Diff.
UI Version
Settings → Personalization → UI Version: switch instantly between v1 (original) and v2 (modern minimalist, default). v1 is preserved as a frozen baseline — flip back any time.
Chat layout
Settings → Personalization → Chat Layout:
Stacked (default) — traditional vertical message flow
Columns — council member responses for each stage render as a side-by-side grid (3-4 cards on desktop, 2 on tablet, 1 on phone). Each card truncates with a "Show full" toggle that expands that card to full width.
Logo + watermark
The full logo is huge on the welcome screen and persists as a barely-visible watermark behind the chat once messages start.
Sidebar
＋ New Session — fresh session at top level
Projects + ＋ New — workspace folders
Sessions — top-level sessions (not in a project)
User badge with role pill
️ Settings
Session row icons
Hover any session in the sidebar to reveal three flat white icons (Lucide-style line glyphs):
Rename — same as /rename slash command. Opens an inline rename field.
Move to project — pick a workspace from the modal (or "Top-level" to remove from any project)
Delete — turns red on hover, asks to confirm
Project settings (gear next to project name) — opens project modal: rename, instructions, files, use-global-context, delete
Post-debate controls bar
Appears at top of chat after a debate. Filter buttons (transcript filter — see Discussion / Filter tab) + ↻ New Round + ️ Send to Arbiter. Persists across refresh.
Chat vs Code mode
Chat (default) — fast streaming responses, no file-system actions.
Code — gives the council direct file-system access on the server so the Final Coder can implement code for you. Skipped in Chat mode.
Toggle in the header. Or type /code / /chat mid-conversation.
Slash commands
/code or /cliSwitch to Code mode (Developer/Admin only)
/chat or /apiSwitch to Chat mode
/imagine [prompt]Generate image with Grok Aurora (default)
/imagine gemini [prompt]Generate image with Gemini Imagen
/rename [new name]Rename current session. The pencil icon does the same.
@ mentions (in discussion mode)
@claude / @chatgpt / @gemini / @grok / @deepseekAddress that member directly — they reply, others stay quiet
Keyboard
EnterSend message
Shift+EnterNew line
Export current session
Download icon in the input row → popup with date range + role filter:
JSON — your full session as a downloadable file
PDF — clean color-coded transcript, noise stripped
Filter by date range and by role (all / AI-only / human-only)
Generated images
Stay inline in the chat across refresh. Right-click → Save As to download.
Native iOS app
There's a native iPhone app (currently via TestFlight). It uses the same iOS-style composer as the web (＋ menu, ↑ send, mode-aware placeholder) and shares your account: your settings, connections, projects, and sessions sync across web and app. Settings are fully editable in-app on iOS too — including My Connections, Personalization, and push prefs — and any change persists back to the web.
Install as a web app (PWA)
iPhone Safari: tap Share → Add to Home Screen. Opens fullscreen with no browser chrome — looks and feels like a native app.
Android Chrome: menu → Install app, or Chrome will offer an Install banner.
Desktop Chrome / Edge: address bar → install icon (or menu → Install LLM Council).
Per-event opt-out checkboxes available. iOS quirk: push only works if you've added the app to the Home Screen (regular Safari can't receive push). Requires iOS 16.4+.
Roles
User — Chat mode only, no LLM Models / Connections / Council / Users / Task Routing settings, model picker shows admin-configured tiers
Developer — Chat + Code modes, otherwise same restrictions as User
Admin — Full access, can switch into other users for support, can broadcast settings to all users
What you can do
Create/rename/delete sessions
Create/rename/delete projects (own projects only)
Move sessions between projects
Edit own user memory + custom global instructions
Manage own connections (Settings → My Connections — Drive / Gmail / Calendar OAuth) — see the Connections tab
Review and approve own Task Blueprints, and view own Audit Log — see the Tasks & Audit tab
Pick UI version (v1/v2), chat layout, push notification prefs
Change own password
Log out (red button at bottom of settings nav, or Account pane)
All of these per-user settings persist across the web app and the native iOS app, and are editable on either.
Admin-only settings
Council — pick which model fills each council seat, Lead Arbiter, Final Coder, Fallback, Briefer. Save and broadcast (megaphone button) pushes settings to all connected users live.
Task Routing — assign best-fit model for each task type (code-gen, code-review, planning, creative, data, summary, math, research, vision, chat).
Connections — API keys per provider, plus per-provider mode (Auto / API / CLI).
LLM Models — Expert / Assistant / Intern model per provider. Refresh Models pulls the latest from each provider.
Users — create/delete users, change role inline. → Switch impersonates any user (banner at top with ↩ Switch Back).
Claude
ChatGPT
Gemini
Grok (xAI)
DeepSeek
