AI Roundtable

One shared conversation for Anthropic Claude Code, OpenAI Codex, and xAI Grok Build — one transcript, three assistants, all running locally on your machine.

🔗 Project page →

Claude, Codex, and Grok working the same problem in one transcript — live activity panes on the left, consensus in the thread.

Install

AI Roundtable requires no third-party Python packages. Install Python, Git, and whichever provider CLIs you want to use. You can enable one provider, any pair, or all three. Python 3.10 or newer is recommended; the automated test suite uses Python 3.12. If GitHub asks during cloning, authenticate with an account that has access to this repository.

Windows

Open PowerShell.

1. Install Python and Git if needed:

   winget install --id Python.Python.3.12 -e
   winget install --id Git.Git -e

2. Install the provider CLIs you want:

   # Claude Code
   irm https://claude.ai/install.ps1 | iex
   
   # OpenAI Codex
   irm https://chatgpt.com/codex/install.ps1 | iex
   
   # Grok Build
   irm https://x.ai/cli/install.ps1 | iex

3. Open a new PowerShell window, clone the repository, and start the app:

   git clone https://github.com/openensemble/AI-Roundtable.git
   Set-Location AI-Roundtable
   py -3 server.py

4. Open http://127.0.0.1:8765/ in your browser.

macOS

Open Terminal.

1. Install Python and Git. With Homebrew:

   brew install python git

2. Install the provider CLIs you want:

   # Claude Code
   curl -fsSL https://claude.ai/install.sh | bash
   
   # OpenAI Codex
   curl -fsSL https://chatgpt.com/codex/install.sh | sh
   
   # Grok Build
   curl -fsSL https://x.ai/cli/install.sh | bash

3. Open a new Terminal window, clone the repository, and start the app:

   git clone https://github.com/openensemble/AI-Roundtable.git
   cd AI-Roundtable
   python3 server.py

4. Open http://127.0.0.1:8765/ in your browser.

Linux

Open a terminal.

1. Install Python and Git using your distribution's package manager:

   # Ubuntu / Debian
   sudo apt update
   sudo apt install -y python3 git
   
   # Fedora
   # sudo dnf install python3 git
   
   # Arch
   # sudo pacman -S python git

2. Install the provider CLIs you want:

   # Claude Code
   curl -fsSL https://claude.ai/install.sh | bash
   
   # OpenAI Codex
   curl -fsSL https://chatgpt.com/codex/install.sh | sh
   
   # Grok Build
   curl -fsSL https://x.ai/cli/install.sh | bash

3. Open a new terminal, clone the repository, and start the app:

   git clone https://github.com/openensemble/AI-Roundtable.git
   cd AI-Roundtable
   python3 server.py

4. Open http://127.0.0.1:8765/ in your browser.

Authenticate the provider CLIs

AI Roundtable reuses the provider CLIs' existing authentication. It does not ask you to copy API keys into the app. Run each installed CLI once and follow its login prompt:

claude
codex
grok

Confirm that the commands are available before starting AI Roundtable:

claude --version
codex --version
grok --version

If a CLI is missing, the app shows an installation notice and disables only that participant. The other installed providers remain usable.

What is AI Roundtable?

AI Roundtable is a local shared conversation between:

• You
• Anthropic Claude Code
• OpenAI Codex
• xAI Grok Build

The assistants do not receive separate, isolated prompts. AI Roundtable owns one shared transcript and gives the selected assistant the complete labeled conversation before each turn. The second assistant can respond to the first, and the third can respond to both.

The browser interface and Python server run locally on 127.0.0.1. Provider responses stream into the page through their installed CLIs, using the authentication those CLIs already maintain.

Quick start

1. Start server.py.
2. Select Claude, Codex, Grok, or any combination using the participant chips.
3. Use Choose… beside the 📁 workspace field to pick the folder the assistants should inspect, or enter its path manually.
4. Leave ✎ Edit off for discussion and inspection, or enable it when you intentionally want the assistants to edit files and run commands.
5. Type a message and press Enter.

By default, every enabled assistant replies in sequence. The first menu chooses who starts; that provider's model and effort controls move into the first position while the other enabled providers retain their order. alternate rotates the starting participant each round. Each provider keeps its own saved model and effort when the order changes.

Address particular assistants

Start a message with a provider's name to choose who responds:

Claude, review this design.
Codex and Grok, compare these two implementations.
@claude @codex @grok, reach a consensus.

An assistant that is not targeted stays silent for that turn but sees the resulting messages in later turns.

Riff, Auto, Stop, and barge-in

• Riff ▸ starts one assistant round without adding a new human message.
• Auto ⟳ keeps the selected assistants taking ordered turns.
• Stop interrupts the active provider process.
• Typing and sending while an assistant is responding barges in, stops that response, and gives the table your new message.
• Esc also interrupts the active response.

Activity panes

The left side shows provider activity while a turn runs:

• streaming reasoning or thinking when the provider exposes it;
• file reads, searches, shell commands, and other tool calls;
• command results and failures.

Grok Build's current streaming format exposes thinking and reply fragments but not every tool event, so its activity pane may contain less detail than Claude's or Codex's.

LAN access

Open AI Roundtable on the server computer and click 📱 LAN off. Enter and confirm a password, then either choose Set password first or choose Enable LAN access to save it and enable access in one step. The server restarts on the local network and shows a normal, copyable LAN address with no password or token embedded in it. Open that address on another device connected to the same network and enter the password. Only a salted, slow password verifier is saved in config.json; the password itself is not. After login, the browser keeps an expiring internal session credential in that tab's sessionStorage and sends it with protected API requests.

Browsers opened directly on the computer running AI Roundtable never need the LAN password, even while LAN access is enabled. Password login applies only to other devices on the network.

LAN mode is authenticated but ordinary HTTP traffic is not encrypted. Use it only on a trusted private network: someone monitoring the network could observe the password or session. Never port-forward the server to the internet. The in-app toggle listens on all IPv4 interfaces (0.0.0.0), including active VPN interfaces, while offering detected private-network addresses as LAN links. Your firewall may ask whether Python can accept private-network connections. WSL 2 users should use mirrored networking when available; reverse proxies and port-proxy forwarding are not supported because local-only controls rely on a direct client connection.

The native folder picker, LAN toggle, and server restart remain restricted to a browser on the computer running AI Roundtable. Chatting, conversation history, model controls, and agent turns work from an authenticated LAN device. The server computer can replace a forgotten password; changing it signs out connected devices. LAN mode survives in-app restarts and resets to off after a fresh launch unless ROUNDTABLE_HOST explicitly requests a network bind.

Workspace access and safety

The 📁 workspace field sets the working directory for every assistant. Choose… opens the host operating system's folder dialog; manual entry remains available for headless or remote sessions. ~ is expanded, and the app checks that the folder exists before starting a provider. The dialog always selects a folder on the computer running server.py and is enabled only when the browser is connecting from that same computer.

Read-only mode

Edit mode is off by default:

• Codex uses its read-only sandbox.
• Grok requests its read-only sandbox where the host supports it.
• Claude uses provider permission controls and an explicit read-only instruction. Native Windows uses Claude's plan permission mode.

Edit mode

Enable ✎ Edit only when you want live changes:

• Claude uses bypassPermissions.
• Codex uses its approvals-and-sandbox bypass flag.
• Grok requests its workspace sandbox and bypass permission mode.

The assistants may edit files, run builds, execute Git commands, or make other changes in the selected workspace. Review their activity and keep important work under version control.

Native Windows note

Codex provides a native Windows sandbox. Claude and Grok run natively but do not currently provide an equivalent Windows OS sandbox. AI Roundtable still applies their permission modes and prompt instructions, but use WSL 2—with the app and CLIs installed inside WSL—when you need stronger isolation for those providers.

AI Roundtable prefers official .exe provider launchers on Windows. A legacy npm .cmd launcher is accepted only when its matching .ps1 shim exists; the included UTF-8 bridge keeps arguments out of cmd.exe. Stop and barge-in use Windows taskkill /T /F to terminate the provider process tree, with parent-only termination as a last resort if Windows denies the tree operation.

Models and effort levels

Each provider has its own model and effort menu. Effort is validated against the selected model rather than shared globally.

• Claude: model capabilities are discovered from Claude Code's initialization response. Models that support the complete effort range expose low through max plus ultracode; Haiku exposes only the controls it supports.
• Codex: models, supported reasoning levels, and model defaults come from Codex's local model cache. A smaller model cannot inherit an unsupported stronger effort such as max.
• Grok: models and reasoning efforts come from Grok Build's model cache. Grok 4.5 currently exposes low, medium, and high.

Unsupported values are never passed through blindly. The server selects a supported lower level when appropriate or omits effort for models that do not expose it.

Model and effort selections are remembered in the browser. Use ⟳ Restart in the header after updating a provider CLI to reload its model and capability catalogs.

Conversations and personal settings

• Conversations are saved as JSON in conversations/.
• The history drawer supports opening, renaming, pinning, and deleting conversations.
• ⭳ exports the current transcript as Markdown.
• Use the 👤 name field and Save to choose the human display name used in bubbles, exports, and provider prompts.
• The display name is stored in config.json. A fresh checkout defaults to You.
• conversations/, config.json, provider-local settings, caches, and temporary files are excluded from Git.

Remember that the entire shared transcript is sent to whichever provider is selected for a turn. Do not include secrets that you do not want that provider to receive.

Running on a different port

Pass a port on the command line:

python3 server.py 9000

On Windows:

py -3 server.py 9000

Then open the matching address, such as http://127.0.0.1:9000/.

Configuration

Variable	Default	Purpose
ROUNDTABLE_HOST	127.0.0.1	HTTP bind address; a non-loopback bind starts LAN mode, but remote APIs remain locked until a password is set locally
ROUNDTABLE_PORT	8765	HTTP port
ROUNDTABLE_CWD	repository folder	Initial assistant workspace
ROUNDTABLE_DATA	conversations/	Saved-conversation directory
ROUNDTABLE_CONFIG	config.json	Local app configuration
ROUNDTABLE_CLAUDE_MODEL	claude-opus-4-8	Default Claude model
ROUNDTABLE_CODEX_MODEL	Codex default	Optional forced Codex model
ROUNDTABLE_GROK_MODEL	grok-4.5	Default Grok model
CODEX_HOME	~/.codex	Codex configuration and model cache
GROK_HOME	~/.grok	Grok configuration and model cache

Troubleshooting

A provider is marked missing

Open a new terminal after installing a CLI, then check:

claude --version
codex --version
grok --version

On Windows, use where.exe claude, where.exe codex, or where.exe grok. On macOS and Linux, use which claude, which codex, or which grok. Restart AI Roundtable after fixing PATH.

A model rejects an effort value

Restart the server so it reloads the provider's current model metadata. AI Roundtable rebuilds each effort menu for its selected model and validates the value again before launching the CLI.

The port is already in use

Choose another port:

python3 server.py 9000

A workspace is rejected

Use Choose… or enter an existing absolute folder path. The green check beside the workspace field shows the resolved folder that will be given to the assistants. If no graphical desktop or native dialog is available, the picker reports an error and manual entry continues to work.

Development and tests

Run the test suite:

python3 -m unittest -v

Windows PowerShell:

py -3 -m unittest -v

GitHub Actions runs the suite on Ubuntu, macOS, and native Windows. The Windows job also installs the current official Codex CLI and launches it through AI Roundtable's process path.

No Python packages or frontend build step are required.