Deliberation and PAL
PAL MCP server (external) (formerly Zen MCP) and Deliberation solve a similar problem: your coding agent asks other models for help over MCP. PAL covers more providers and has more tools. Deliberation is smaller and focuses on cross-model review, with rules that keep the orchestrating agent from approving its own plan.
Already on PAL? The migration guide maps tools and settings.
PAL at a glance
As of 2026-11-17, from the GitHub API and PAL's README:
- 11,842 GitHub stars, Apache-2.0 (the license in its
LICENSEfile). - Latest release v9.8.2, published 2025-12-15. The last push to the repository was also on 2025-12-15.
- Python 3.10+. Install with
git cloneand./run-server.sh, or withuvxin an MCP config. - Providers: Gemini, OpenAI, Azure OpenAI, X.AI (Grok), OpenRouter, DIAL, Ollama and other
custom OpenAI-compatible endpoints. A provider turns on when its key is in
.env. - Tools enabled by default:
clink,chat,thinkdeep,planner,consensus,codereview,precommit,debug,apilookup,challenge. Disabled by default to save context:analyze,refactor,testgen,secaudit,docgen,tracer.
Where PAL is the better fit
- Provider breadth. Azure OpenAI and DIAL have no Deliberation equivalent. PAL calls the
OpenAI API with a key; Deliberation reaches GPT only through the
codexCLI. - Code from capable models. PAL's
chatcan return complete code from capable models. Deliberation's experts are prompted to advise; only Gemini, through the Claude Code plugin, can edit files itself. - Tool range.
clinkstarts other AI CLIs (Gemini CLI, Codex CLI, Claude Code) as subagents with their own context.precommit,refactor,testgen,docgen,tracer,apilookupandchallengehave no dedicated Deliberation tool. - Vision. PAL sends images to models that accept them. Deliberation is text only.
- Long threads. PAL carries a conversation across tools and models, and can revive Claude's context after a reset ("continue with O3"). Deliberation keeps follow-up threads only for Gemini, Grok and OpenRouter in the Claude Code plugin, and they are lost on restart.
- Large prompts. PAL works around the MCP prompt size limit. Deliberation relies on each provider's context window.
- Adoption. PAL has 11,842 GitHub stars and Deliberation has 184 (both as of 2026-11-17).
Where Deliberation is the better fit
-
Consensus that guards against the orchestrator. In
/consensus, the agent that runs the loop also votes and decides which objections count, so it could pass its own plan. These rules stop that:- The agent records its verdict before the panel answers, and the engine holds the panel back until it does.
- A round converges only when at least
consensus.quorumFloormodels answer (default 2, capped at the panel size), every one of them approves, no accepted critical issue is left, and the agent approves too. - Every dismissed issue needs a written reason, and the engine rejects a dismissal without one.
The loop stops at
consensus.maxRounds(default 5). PAL'sconsensusassigns stances (for, against, neutral) and the agent combines the answers. -
Disagreement you can read.
/ask-allreturns each model's answer separately and the synthesis lists where they agree and where they split. -
A CLI login for GPT. GPT runs through
codex exec, so those calls use your Codex login and plan instead of an OpenAI API key. Gemini uses aGEMINI_API_KEYby default; the Antigravity CLI (agy) is an opt-in path (read the terms note first). -
Live config. Model records, routing and the OpenRouter block in
~/.config/deliberation/config.jsontake effect without a restart. Each record says whether it joinsask-allandconsensus, and which experts it serves. -
Packaging. A Claude Code plugin from a marketplace, an npm package (
@antonbabenko/deliberation-mcp, Node 18+), and native plugins for Codex, Cursor, Kiro, Antigravity, Gemini CLI and OpenCode. -
Measuring the panel. An opt-in debug log, saved runs, the
analyzetool (latency, tokens and verdict agreement per model) and a local read-only dashboard help you drop models that add little.
Side by side
| PAL | Deliberation | |
|---|---|---|
| License | Apache-2.0 | MIT |
| Language | Python | JavaScript, no runtime dependencies |
| Install | git clone + ./run-server.sh, or uvx |
Claude Code plugin, npx -y @antonbabenko/deliberation-mcp, native host plugins |
| GPT | OpenAI API key | Codex CLI login |
| Gemini | Gemini API key | GEMINI_API_KEY (AI Studio); Antigravity CLI (agy) opt-in |
| Grok | XAI_API_KEY |
XAI_API_KEY |
| OpenRouter | OPENROUTER_API_KEY, models in conf/openrouter_models.json |
OPENROUTER_API_KEY, models as records in config.json |
| Local models | Custom endpoint (Ollama, vLLM, LM Studio) | Ollama and LM Studio providers; other endpoints through providers.openrouter.apiBase |
| Azure OpenAI, DIAL | Yes | No |
| Vision | Yes | No |
| Code | Any capable model returns code for your agent to apply | Any provider can return code as text; only Gemini (Claude Code plugin) edits files itself |
| Consensus | Stances per model, agent combines | Blind verdict first, no self-approval, reasons for every dismissal, round cap |
| Parallel answers | Through consensus |
ask-all |
| Personas | One system prompt per tool | Eight experts usable with any provider |
| Pre-commit check | precommit |
Run code-reviewer or /consensus on the diff |
| CLI subagents | clink |
No |
| Config changes | Edit .env, restart |
Model records, routing and the OpenRouter block reload live |
Using both
The two servers register under different server names, so you can keep both. One way to split
the work: PAL for clink, precommit, vision and Azure, and Deliberation for /consensus and
/ask-all reviews.