Agents
A Palabre agent is a configuration entry that describes how to call an AI.
Palabre supports two agent families:
- CLI agents, such as Claude Code, Codex CLI, Antigravity CLI, OpenCode, or Mistral Vibe;
- local agents, such as Ollama via its HTTP API.
CLI agents
A CLI agent is launched as a terminal command. Palabre sends it a prompt, retrieves its output, then closes the process.
Each CLI keeps its own rules: authentication, default model, subscription, usage limits, and configuration options. Some CLIs can work with a free tier, a subscription, a local provider, or a cloud provider; in all cases, quotas and rate limit errors come from the agent being used, not from Palabre.
If a CLI exits with an error without writing to stderr, Palabre shows a sanitized, bounded stdout excerpt to help diagnose the failure, for example an authentication error. The session still fails: this excerpt is never accepted as a response. Fix authentication or configuration in the affected CLI first, then run Palabre again.
Local agents
A local agent like Ollama uses a model installed on your machine. This is useful for critic, exploration, or lightweight summary roles. If you configure Ollama with a cloud offer or a paid service, the associated limits remain those of Ollama or the chosen provider.
Ollama does not read your files directly. To give it context, use --files or --context.
View available agents
palabre agents
palabre agents --json
If an installed CLI does not appear in the configuration:
palabre config --sync-agents
Roles
Each agent has a role that influences the prompt:
| Role | Usage |
|---|---|
implementer | Propose a concrete solution. |
reviewer | Look for risks, regressions, and missing tests. |
critic | Challenge assumptions. |
architect | Structure technical options. |
scout | Quickly explore a topic. |
summarizer | Produce a faithful summary. |
Roles do not change the model used. They change the instruction sent in the prompt.
Using a role
A persistent role is configured in palabre.config.json, in the agent entry, or with palabre agent-role <agent> <role>. For a one-off run, use --role-a, --role-b in debate mode, or --ask-role to apply one shared role to Ask agents.
Example:
Excerpt from the agents block:
"codex": {
"type": "cli",
"command": "codex",
"role": "implementer"
},
"claude": {
"type": "cli",
"command": "claude.exe",
"role": "reviewer"
}
With this configuration:
palabre codex-claude "Critique this plan" -t 4
Codex receives a concrete proposal instruction, while Claude receives a critical review instruction.
To use two different behaviors with the same CLI, create two distinct agents in the configuration, for example claude-reviewer and claude-summarizer.
Models passed as arguments
When you use --model-a, --model-b, or --summary-model, Palabre does not verify the list of models available at the provider. The value is passed to the relevant CLI via its configured model option. If the model does not exist or is not authorized by your account, the CLI will return the error.
Agents and Relay
Relay targets Codex or Claude Code conversations: closed by default; --open targets only Codex on Windows, macOS, or Linux (pilot, with python3 required on POSIX). Mission, a Relay option, uses closed conversations. Antigravity, OpenCode, Mistral Vibe and Ollama participate in Debate, Ask and Chat, not as Relay targets.
CLI agents run from the current directory. Defaults request read-only tools where available: Claude Read,Glob,Grep, Codex --sandbox read-only, Vibe read,grep. OpenCode keeps its usual environment: configured plugins, MCP servers and permissions stay active, with no read-only access or neutralization promised. Custom arguments remain under your control; migrations replace only exact old defaults. A recognized limit gives usage-limit, without automatic resend.
Windows
Adapters prefer a native executable, then a recognized npm or pnpm shim launched directly (Node + JavaScript script, or the native npm variant). For pnpm, the shim's NODE_PATH is passed through. Other shims may use PowerShell; remaining shell wrappers are refused for an argument prompt. shell: true in generated configuration enables this conservative resolution; do not remove it to force a wrapper launch. Relay accepts only direct executables and the npm Node + script form.
tier (primary or local) classifies the configured agent; it does not change its role instruction.