AgentCouncil
Multi-agent deliberation protocols for AI coding assistants.
Newest in 0.6.1: native Cursor plugin support — install AgentCouncil directly from the repo URL in Cursor (0.6.0 added the /configure skill and a self-healing server bootstrap).
AgentCouncil now runs natively in Cursor — alongside Claude Code and Codex — via a Cursor MCP config and generated slash commands. The default outside backend follows the host you run on (Claude Code → Claude, Codex → Codex, Cursor → Cursor), so deliberations work out of the box on every host, and on Cursor you can point the outside agent at any Cursor model (e.g. gpt-5 vs sonnet-4.5).
Underneath that headline feature, AgentCouncil still provides its core multi-agent protocols: /brainstorm, /review, /decide, /challenge, and /inspect. The host agent — Claude Code, Codex, or Cursor — works with an outside agent that stays independent by role, backend, and session. The outside agent defaults to the host you run on (Claude Code → Claude, Codex → Codex, Cursor → Cursor), or you can configure Cursor models, Codex, Ollama, OpenRouter, Bedrock, or Kiro for cross-model diversity. See CURSOR.md for running on Cursor.
What's New In 0.6.0
This release adds a self-configuration skill and hardens the server bootstrap:
/configure skill: set up AgentCouncil from inside the agent — it reads your current config, detects which backends are available, and writes a valid .agentcouncil.json (profiles + default outside/lead backend) for Cursor models, Codex, Claude, Ollama, OpenRouter, or Bedrock. API keys are never written to the file (api_key_env holds the env var name).- Self-healing bootstrap:
scripts/start-server.sh now verifies and repairs a partial/corrupt dependency install instead of crashing the MCP server with No module named 'rich' — fixing intermittent startup failures (notably the first launch inside Cursor).
Previously in 0.5.0
This release brings AgentCouncil to Cursor and makes the default backend host-aware:
- Runs natively in Cursor: ships
.cursor/mcp.json and generated .cursor/commands/*.md slash commands (one per skill), so /brainstorm, /review, /decide, /challenge, /inspect, and /autopilot work inside Cursor. See CURSOR.md.
- Host-aware default backend: the default outside agent (and library-mode lead) is now the backend it runs on — Claude Code →
claude, Codex → codex, Cursor → cursor — falling back to claude when no host is identified. Explicit backend=, env vars, and default_profile still win. (Behavior change for Codex hosts — see CHANGELOG.md.)
- Cursor model selection: a new
cursor backend runs the cursor-agent CLI; name profiles like cursor-gpt5 / cursor-sonnet to use different Cursor models per deliberation, independent of the editor's model.
Previously in 0.4.0
Lead-agent selection became first-class in MCP/library mode:
- Configurable lead agent:
brainstorm, review, decide, challenge, review_loop, and protocol_resume accept lead_backend= and lead_model=.
- Native lead adapters:
lead_backend may be claude, codex, or a named profile backed by one of those native CLI providers. Claude still defaults to opus; Codex uses the Codex CLI default unless configured.
- Independent outside selection:
backend continues to select the outside agent. Same-backend pairings are allowed with separate sessions and recorded as same_backend_fresh_session.
- Autopilot gate wiring: lower-level MCP autopilot runs persist
lead_backend/lead_model and use them for real gate execution when AGENTCOUNCIL_AUTOPILOT_GATES=1.
- Native Codex plugin surface:
.codex-plugin/plugin.json and .mcp.json let Codex load AgentCouncil skills and MCP tools directly. Skill wording is host-neutral so Codex and Claude Code can both be the lead in skill mode.
- Release hardening: explicit unknown backend/profile names now fail closed, real gate transcripts carry lead/outside provenance, real review/challenge/review_loop gates run certification checks, and partial review failures escalate instead of passing.
The 0.3.x series added the current /autopilot workflow foundation:
