Roundtable
Fans out a question to multiple frontier models in parallel, optionally deliberates to consensus, and returns one decided answer from the command line or any MCP-capable coding harness.
README
<div align="center">
🪑 Roundtable
A council of frontier models — for the decisions that matter.
One question → many minds → one decision.
Fan a question out to Grok, Codex/OpenAI, GLM, MiniMax, Claude, and Gemini in parallel, optionally deliberate to consensus with a chaired vote, and get back one decided answer — from the command line or any MCP-capable coding harness.
</div>
curl -fsSL https://roundtable.sh/install.sh | bash
Why Roundtable?
Every model has a blind spot it can't see. Right before an irreversible call — an architecture, a migration, a rewrite — is exactly when one model's confident answer deserves a second (and third, and fourth) opinion.
Roundtable convenes a panel of frontier models. Each answers independently; then, for
high-stakes calls, a multi-round deliberation lets every head read the whole table and
revise until a Claude chair declares CONSENSUS or names the remaining disagreement.
Disagreement is signal, not noise.
- 🗳️ Advisory (default) — every head answers once, blind to the others. Raw, independent reads.
- ⚖️ Deliberation (
--rounds N) — heads see each other and revise; a neutral chair rulesCONSENSUS/CONTINUEand stops early on agreement. - 🔑 API keys are enough — no local CLIs required. Missing heads are skipped; the council runs on whatever subset you've configured.
- 📦 One binary, everywhere — a CLI, an MCP server, and a one-command installer that wires itself into your coding harness.
Example
$ roundtable "Reply with exactly: OK three words" --heads claude,gemini
# 🪑 Round Table — 2 heads — advisory
## 🔵 Claude (claude-opus-4-8, Anthropic) → OK three words
## ✨ Gemini (gemini-3.1-pro-preview, Google) → OK three words
Add --rounds 3 and the heads deliberate; a chair prints VERDICT: CONSENSUS (or
CONTINUE) with the points they agreed on and any dissent named.
Install
curl -fsSL https://roundtable.sh/install.sh | bash
Or pick your package manager — all four deliver the same engine:
pip install roundtable
uv tool install roundtable
brew install frontier-infra/tap/roundtable
The installer places the engine under ${XDG_DATA_HOME:-~/.local/share}/roundtable,
symlinks roundtable into ~/.local/bin, and prints a PATH hint if needed.
Quickstart
# 1. Configure keys (masked prompts; shows which heads are ✓/✗). Stored chmod-600
# in ~/.config/roundtable/config.env. Any subset works — missing heads are skipped.
roundtable auth
# 2. Ask the council
roundtable "Postgres or SQLite for a single-tenant internal tool?"
# 3. Deliberate to consensus on a high-stakes call
roundtable ask -q "Should we rewrite the billing service in Go?" --rounds 3
roundtable doctor shows which heads are configured and pings each for reachability.
Flags
-q/--question · -c/--context <file> (shared background for every head) ·
--heads grok,codex,openai,glm,minimax,claude,gemini · --rounds N (deliberation) ·
--research (web on, multi-step) · --timeout <secs> · --out <file>.
Exit 0 if at least one head answered.
The heads
| Head | Provider | Key / auth |
|---|---|---|
| 🤖 Grok | xAI | XAI_API_KEY (or local grok CLI / Hermes OAuth) |
| 🧠 Codex | OpenAI | local codex CLI (read-only) |
| 🧠 OpenAI | OpenAI | OPENAI_API_KEY (direct API — no local CLI needed) |
| 🟣 GLM | Z.AI | ZAI_API_KEY |
| 🟠 MiniMax | MiniMax | MINIMAX_API_KEY |
| 🔵 Claude | Anthropic | ANTHROPIC_API_KEY (also the deliberation chair) |
| ✨ Gemini | GEMINI_API_KEY / GOOGLE_API_KEY |
Heads with no key/CLI are skipped inline — the council runs on whatever subset you've configured, so API keys alone are enough (no local CLIs required).
MCP + wire into your harness
Roundtable ships an MCP server as a subcommand of the one binary:
roundtable mcp serve # stdio JSON-RPC; one tool: roundtable(question, heads?, rounds?, research?)
roundtable mcp config # prints the server JSON block to paste manually
Auto-wire detected harnesses (idempotent — backs up first, never clobbers):
roundtable install # detect + wire everything present
roundtable install --harness cursor,codex
roundtable install --all # non-interactive
Wires Claude Code (skill + MCP), Cursor (~/.cursor/mcp.json),
Codex CLI (~/.codex/config.toml), and prints a generic JSON block for anything else:
{ "command": "roundtable", "args": ["mcp", "serve"] }
Use in Claude Code
Roundtable ships as a Claude Code plugin — a roundtable skill that drives the CLI.
Two ways to add it:
A — Plugin marketplace (inside Claude Code):
/plugin marketplace add frontier-infra/roundtable
/plugin install roundtable@roundtable
The skill calls the globally-installed roundtable CLI, so install that first:
curl -fsSL https://roundtable.sh/install.sh | bash.
B — roundtable install — auto-wires the skill and the MCP server into Claude
Code (plus Cursor / Codex) in one step:
roundtable install
Commands
roundtable "question" shorthand for `ask`
roundtable ask run the council (all engine flags)
roundtable auth [provider] configure API keys (masked, chmod-600)
roundtable models per-head model picker
roundtable heads enable/disable heads
roundtable doctor which heads are configured + reachability ping
roundtable config show / edit config.env
roundtable mcp serve|config MCP stdio server / print config
roundtable install auto-wire detected harnesses
roundtable version | update show version / upgrade in place
For agents
Building on top of Roundtable? See AGENTS.md — a concise guide for an
AI agent to install, configure, run, and wire Roundtable into its own harness.
Links
- Website & docs — https://roundtable.sh
- Source — https://github.com/frontier-infra/roundtable
- Frontier Infra — https://frontierinfra.org
License
MIT © Frontier Infra
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。