antonbabenko.com

Deliberation has a website: usedeliberation.com

Deliberation now has its own site with recorded consensus runs, comparison pages, install guides for each host and search across the docs.

Three lines in navy, teal and light blue curve into one orange dot on a pale blue background, and a dashed line splits off below them

Six months after the first post, Deliberation has a home page: usedeliberation.com. It went live today. Deliberation is an open-source MCP server and agent plugin that sends your coding agent's plan or diff to models from other vendors for a second opinion. The site has recorded runs, comparison pages and an install guide for each host, so start there.

Until now everything lived in the GitHub repository: the README, the docs folder and the changelog. That works for people who already found the repository. It does not work for someone who heard about Deliberation in a talk, or for a colleague I send a link to who never opens GitHub. They now get a site with docs, search and install steps for their own host.

Deliberation works with the agent you already use (Claude Code, Codex, Cursor, Kiro, Gemini CLI, Antigravity, OpenCode), and you get back one verdict with the disagreement kept visible. I use it every day for the work on compliance.tf, same as before.

The home page plays back a real run

The first thing on the home page, after the install line, is a consensus run recorded today on Deliberation 4.6.1. A plan-reviewer panel of four models (Grok, GLM, GPT and Gemini) reviewed a plan to refuse provider calls when the config file exists but cannot be read. In round 1 Grok and GPT asked for changes and the host agent accepted all 10 critical issues in the pool. The plan was revised, and in round 2 all four approved. Click any box and you read that reviewer's answer in full.

I wanted this on the site before any feature list: one real question, the exact prompt, each answer verbatim and the rulings in between.

The sample runs page has nine of them, from plan reviews to Terraform questions and a security review of the dashboard. Six of the nine ended unresolved. I kept them, because a panel that cannot agree is telling you something about the question, and that is where I would not trust one model's confident answer.

Two of the runs are older: a review of a CloudFront and API Gateway Terraform setup from September 5, recovered from saved records and logs. Some parts of those runs were never saved, such as the host's blind verdict and some reviewers' votes. The page marks each of them as "not recorded" instead of guessing. The Terraform code they reviewed sits in a private demo repository, so those two runs have no link to it.

Comparison pages answer the question I get

The question I hear after a talk is some version of "how is this different from X?". The site now has a page for each X I have been asked about:

  • the Codex plugin, OpenAI's codex-plugin-cc for Claude Code. It runs Codex reviews and can hand work to Codex. The two plugins do different jobs, they share one Codex login, and the page shows how to install both.
  • PAL (formerly Zen MCP). PAL covers more providers and has more tools. There is also a migration guide that maps its tools and settings.
  • Council of High Intelligence, a structured debate between 18 analytical personas.

Each page starts with the facts about the other project and the date they were checked (2026-11-17 for all three), because star counts and release dates change. The Codex plugin and PAL pages also say where the other project is the better fit. If a comparison page only lists wins, I stop reading it.

Docs, search and an install for each host

The rest is the docs from the repository, built into pages:

  • one install guide per host, with the exact command for Claude Code, Codex, Cursor, Kiro, Gemini CLI, Antigravity, OpenCode and any other MCP host;
  • the features page with the commands, tools and settings;
  • the eight experts, the providers, the reference pages and the changelog;
  • search across all of it, built with Pagefind when the site builds, so a search loads nothing from another server. Press / or Cmd/Ctrl-K to open it.

The pages that come from the repository are generated on each build, so they follow the code. The pages I write by hand (features, install copy, talks) have a checklist in the repository's AGENTS.md that asks for the site update in the same pull request as the feature change.

The privacy page lists what Deliberation sends to each provider and what it keeps on your machine. Deliberation itself has no telemetry.

Where I will talk about it next

There is a talks page with the slides from "One model is a guess" at AWS Community Day Poland 2026. It walks one Terraform plan through four scenarios that every scanner passes, and shows how a panel reviews the same plan.

At AWS re:Invent 2026 I have a chalk talk, COM302 "The judgment gap: when Terraform scans clean but fails the audit", on Monday, November 30 at 1:30 pm, and again on Thursday, December 3 at 2:30 pm, both in Caesars Forum, Academy 414. Deliberation will come up there too. Reserve a seat in the AWS re:Invent app.

Try it

The fastest start on any host is the first-run command. It finds your API keys and CLI logins, writes the config, runs a demo consensus and prints the install line for your host:

Terminal window
npx -y @antonbabenko/deliberation-mcp init

In Claude Code the plugin install is still three lines:

/plugin marketplace add antonbabenko/agent-plugins
/plugin install deliberation@antonbabenko
/deliberation:setup

Open usedeliberation.com, read a run or two, and tell me what is missing or wrong on the site (good and bad). Issues and pull requests go to github.com/antonbabenko/deliberation.