Human-in-the-loop · auditable

Mediator Agent

An impartial AI mediator for product-manager vs. architect disputes. It weighs both sides against your recorded criteria and recommends a decision — you approve it.

View on GitHub Quick Start Docs
Vendor-independent OS-independent No Azure Bot needed MIT licensed

Run it anywhere

CLI

Try a mediation locally in two minutes — no chat platform required.

Any agentic CLI

An MCP server drives it from Claude Code, Codex, Cursor, Windsurf, …

Slack

A Socket-Mode bot mediates directly in a channel.

Microsoft Teams

Via a Workflow webhook — no Azure Bot resource — or the Bot Framework adapter.

Generic webhook

Plain JSON in, replies out. Bridge Discord, Mattermost, cron, or your own app.

Your choice of model

Claude, OpenAI, or any OpenAI-compatible endpoint (Azure OpenAI, Ollama, LM Studio, …).

The protocol

  1. ExplorationSeparate positions from the interests underneath them; ask targeted questions.
  2. FactsSurface missing or disputed facts — effort, deadline, tech debt, A-SPICE impact.
  3. SteelmanRestate both sides in their strongest form and get the parties' confirmation.
  4. DecisionA fixed, auditable format: DECISION / RATIONALE / CONDITIONS / DISPUTED FACTS / ESCALATION.

A code-level guard blocks a final decision until every party has spoken at least twice — unless they explicitly ask, in which case the uncertainty stays flagged.

Get started

pip install -e ".[all]"
cp .env.example .env        # add your ANTHROPIC_API_KEY or OPENAI_API_KEY
python -m mediator.cli