Plug and Play Reviewer
Plug and Play ReviewerPlug and Play Reviewer

Agent surfaces

Four ways for another agent to start a review. They share one core.

Model Context Protocol

A JSON-RPC tool server. tools/list lists three tools. tools/call runs one.

  • review_pull_request { owner, repository, pull_request }
  • list_findings { review_id }
  • list_remediation_prompts { review_id }

Machine-readable command line

The same three operations as reviewer agent-json subcommands. One JSON object per line on stdout.

  • reviewer agent-json review --owner OWNER --repository REPO --pull-request N
  • reviewer agent-json findings --review-id ID
  • reviewer agent-json remediation-prompts --review-id ID

Agent Client Protocol

The same three actions for an ACP client.

  • review_pull_request { owner, repository, pull_request }
  • list_findings { review_id }
  • list_remediation_prompts { review_id }

Agent2Agent

An agent card advertises one skill, pr-review, over JSON-RPC.

  • agent card: protocolVersion 0.3.0, preferredTransport JSONRPC
  • skill: pr-review (start a review, or fetch findings and prompts)

If GitHub is not connected, every surface returns a typed refusal. It never guesses and never runs a partial review.