Debate
"I want two positions to confront each other."
Two agents answer in turn; a summary can set out agreements and disagreements.
01 / PALABRE
Let your AI tools work together. Keep the decision.
Compare their analyses, have your changes reviewed and move your work and decisions forward. Palabre uses your existing tools and subscriptions; you choose the mode that fits.
npm install -g palabre02 / RELAY
Coordinate work between your agents
From your usual conversation, your tool becomes the lead. With Relay mode, it sends a task or review request to a teammate, in a new conversation or one you choose, then brings the reply back. You no longer have to copy messages between windows.
- Agent A
- PALABRE
- Agent B
- reply
- you decide
Fictional example: lead agent and teammate
You prepare a change with Claude Code. It asks Codex for a review through Relay, with the useful files and context. Claude examines the findings, presents disagreements and proposes corrections. You decide what happens next. Codex can also lead, with Claude as its teammate.
What stays under your control
- Separate conversations. The lead passes the useful context. Relay does not automatically share tools' memories or quotas.
- Permissions depend on the path. The message grants no additional rights. In the background, Palabre limits Codex and Claude Code tools; in an already open Codex conversation, that conversation's permissions apply.
- An explicit next step. No autonomous loop or automatic resend. A
.relay.mdexport lets you reread the exchange; receiving a reply does not prove task success.
Relay connects Claude Code and Codex both ways. Codex can receive in an already open conversation on Windows, macOS and Linux (python3 required on macOS/Linux); immediate display is not guaranteed. Sending to a Claude Code conversation that is in use is not yet supported. Requirements, permissions and limits.
03 / DEBATE
Confront opinions before you decide
With Debate mode, two agents answer the same question in turn. Each one reads the other's answer, challenges or completes it. When enabled, the final summary sets out agreements, disagreements and proposed actions.
- Topic
- Agent A
- Agent B
- summary
- you decide
Fictional example
You are unsure whether to split a module that has grown too large. Codex, as implementer, proposes a split into three parts; Claude Code, as reviewer, spots two hidden dependencies. Codex adjusts. The summary keeps the split, flags the remaining risk and suggests a first test to write. You decide, with the arguments in hand.
- Two perspectives to cross-check. Roles steer instructions, not permissions. Agreement between agents does not prove a proposal is correct.
- A useful summary. Consensus, disagreements and uncertainties, proposed actions, conclusion.
- The right pace. From 1 to 20 responses, with the option to stop earlier on an explicit agreement.
04 / MODES
Which mode for what?
Palabre offers four modes: Debate, Relay, Ask and Chat.
Relay
"I want my agents to cooperate and hand off to each other, without any copy-paste from me."
One agent sends a message directly to another agent (Claude Code or Codex), which replies. You are no longer the go-between, but the decision-maker.
Ask
"I want several independent opinions."
1 to 4 agents answer without seeing each other's answers; a summary can compare them.
Chat
"I'm chatting with an agent, and sometimes need a second opinion to double-check an answer."
Ask another agent with /consult, without leaving the conversation.
In doubt: Debate for a decision with stakes, Relay to coordinate work across several agents, Ask for several opinions, Chat to talk with one agent.
Debate, Ask and Chat start new exchanges; unlike Relay, they do not pick up your existing Codex or Claude Code conversations.
05 / AGENTS
Which tools work with which modes?
Codex, Claude Code, OpenCode and the other tools use your accounts and models. Palabre calls the assistants you bring into the work agents. Each tool must be installed and configured; its presence does not guarantee every mode works.
| Tool | Relay | Debate, Ask and Chat |
|---|---|---|
| Codex | Yes | Yes |
| Claude Code | Yes, in the background | Yes |
| OpenCode | Copy pilot (1) | Yes |
| Antigravity CLI | No | Yes |
| Mistral Vibe | No | Yes |
| Ollama | No | Yes |
(1) Experimental OpenCode Relay pilot: Windows, CLI 2.0.32 only; Desktop not verified. Usual OpenCode environment, with no enforced read-only mode. Tools and requirements.
06 / PRIVACY
The Palabre CLI does not collect your data.
No Palabre account, no CLI telemetry, no Palabre server receiving your conversations. Exports are saved on your machine.
Remote tools and models receive the prompts and context they need, under their own policies and your settings. Ollama does not always mean local: a cloud model is still remote, even when accessed through a local Ollama server.
To keep AI exchanges on your machine, launch Debate, Ask or Chat from your terminal, without a remote assistant, with a local Ollama server and local models for all participants and summaries. Privacy and limits.
07 / GET STARTED
Get started
From your assistant.
Install the Palabre skill in Codex or Claude Code, then reload it. Your assistant needs permission to run local commands.
Ask in plain language.
“Use Palabre Relay to ask Claude Code to review this plan without changing files.” Your lead runs Palabre and brings the result back. Your first Relay.
From a terminal.
palabre opens interactive menus for Debate, Ask and Chat. palabre doctor checks your configuration; direct commands also let you use Relay.
Choose the context.
For Debate, Ask and Chat, --files and --context attach useful files or folders. For Relay, the lead prepares the request and its links.
Give a role.
Roles complement Debate, Ask and Chat: propose, challenge, review. They steer instructions, not permissions, and do not configure Relay.
Keep a record.
Exports (.debate.md, .relay.md, .ask.md, .chat.md) are saved in .palabre/ by default. Check for sensitive information before sharing them.
08 / WHERE TO USE IT
Where you already work
Terminal (CLI)
the palabre CLI, with its interactive interface or commands to use in your scripts.
CLI agents
with the Palabre skill, Codex and Claude Code can run the mode that fits your request from the terminal.
Desktop apps
in a compatible app such as Codex or Claude Code, ask your agent to use Palabre. Type or dictate your request if the app supports voice input; voice comes from the app, not Palabre.
VS Code
a sidebar to start a Debate, Chat or Ask and follow the session live.
One project, several agents, your decisions.
Install PALABRE, pick your mode, coordinate your agents' work and keep the decision.