Use multiple AI agents from the terminal
Palabre is a multi-agent CLI that orchestrates the AI tools already installed on your machine. It can make two agents discuss a decision, compare several independent answers, or let you work with one agent while explicitly consulting a second opinion.
Palabre is not an additional provider. Each agent keeps its own authentication, model, quotas, and access rules.
Choose the right multi-agent mode
| Mode | Agents | Interaction | Primary use |
|---|---|---|---|
| Debate | 2 | Agents answer each other | Challenge a decision and resolve disagreements |
| Relay | 1 target conversation | One message, one reply | Hand off without copy-paste |
| Ask | 1 to 4 | Independent answers | Compare analyses without mutual influence |
| Chat | 1 active, explicit consultations | Ongoing conversation | Work with one agent and request a second opinion when needed |
Make two agents debate
A debate injects the previous transcript into every new turn:
palabre codex-claude "Review the risks in this architecture" -t 4
Presets choose a pair and speaking order. They do not automatically change configured models or roles.
See Run a Codex and Claude Code debate for a complete workflow with roles and project context.
Compare several answers with Ask
Ask sends the same subject to several agents without sharing their intermediate answers:
palabre ask "Compare these migration strategies" --agents codex claude opencode
An optional final summary can then present agreement and differences. Ask is particularly useful when you want to prevent the first response from influencing the next ones.
Chat and consult another agent
Open the TUI, then switch to Chat:
palabre
/chat
/agents codex
Use /consult <agent> when you need a complementary opinion. Chat keeps bounded recent context and exports the session with /end.
Add agents
Palabre detects known CLIs and safely synchronizes the configuration:
palabre config --sync-agents
palabre agents --json
palabre presets --json
Supported agents include Codex, Claude Code, Antigravity, OpenCode, Mistral Vibe, and Ollama.
Give agents project context
Add exact files with --files or scan a folder with --context:
palabre codex-claude "Review this implementation" --context src tests -t 4
CLI agents may also have their own workspace access depending on their tools. Ollama only sees context explicitly injected by Palabre.
Keep an actionable record
Exported sessions write Markdown to outputDir (.palabre/ by default):
.debate.mdfor Debate;.relay.mdfor Relay;.ask.mdfor Ask;.chat.mdfor Chat.
These files preserve the subject, agents, transcript, and optional summary. They provide a shareable record and can be opened by the VS Code extension at the end of a session; it follows NDJSON for events.
Continue with Choose a mode, Context and files, or the agent overview.
Hand off with Relay
palabre relay --from claude:<sender-session> --to codex:<target-session> "Read the plan and list the risks"
The target is closed by default. --open targets open Codex on Windows, macOS, or Linux (pilot; python3 required on POSIX), keeping its own tools and permissions; immediate display is not guaranteed. Relay's Mission option keeps several read-only steps within an agreed scope and budget. See Relay.
Agents: OpenCode, Mistral Vibe, Ollama.