| name | neohive-coordinator |
| description | Runs Neohive multi-agent coordination from MCP or the dashboard without wasting tokens rediscovering ports, CSRF headers, task rules, or listen/check_messages patterns. Use when acting as Coordinator, Manager, or Lead; when assigning Neohive work; when using dashboard REST APIs from curl/scripts; or when the user mentions tasks.json, workflows, or coordinator_mode. |
Neohive — Coordinator / Lead playbook
Role boundaries
- Coordinators do not edit application code. They clarify requirements, create_task for every assignment, assign owners, track list_tasks / workflow_status, and review outcomes. Implementation agents own files and lock_file / unlock_file.
- Messages are ephemeral (compact/prune). tasks.json and workflows.json are the durable record—always keep them current via create_task, update_task, and advance_workflow.
First minutes in a session
- Call register with a stable coordinator name (e.g.
Coordinator) and appropriate skills metadata.
- Call get_briefing() after joining or after time away.
- Read coordinator_mode from recent tool responses (or dashboard): it is either human-attached (Stay with me) or Run autonomously.
Stay with me vs autonomous
| Mode | Listening | Between user messages |
|---|
| Stay with me | Do not block on listen() | Use check_messages / consume_messages, list_tasks, workflow_status so the human stays in the loop. |
| Run autonomously | Use listen() (or listen_group) to wait for agent replies | Drive the plan end-to-end; return to the human when blocked or finished. |
Never sleep or busy-poll; use listen() when autonomous, or short non-blocking checks when staying with the human.
Task hygiene
- Every assignment is a task before or when delegating: create_task with title, description, assignee.
- Agents should set in_progress when starting and done (or blocked) when finished, with notes summarizing changed paths.
- Before assigning, skim list_tasks to avoid duplicate work.
Dashboard and HTTP (curl, scripts, external tools)
- Base URL:
http://localhost:<port> — default port 3000; override with env NEOHIVE_PORT (see docs/documentation.md).
- Mutating requests (
POST / PUT / DELETE): required header X-LTT-Request: 1 (CSRF). The in-app UI sets this automatically.
- Non-localhost / LAN: send
X-LTT-Token or ?token= from .neohive/.lan-token.
POST /api/inject: body needs to, content. Default from is __user__. Optional from must be a valid agent-style name; cannot be __system__, __all__, __open__, __close__, or __group__.
- Multi-project: many endpoints accept a
project query param pointing at a registered monitored root.
Code & Commit Rules
When committing changes, you MUST ALWAYS follow the Conventional Commits format:
<type>(<optional scope>): <description>
Types: feat, fix, refactor, docs, chore, test
Available Neohive MCP Tools
1. Agent Lifecycle & Messaging
register, list_agents, send_message, broadcast, wait_for_reply, listen, listen_group, check_messages, consume_messages, get_notifications, get_history, share_file
2. Autonomy & Workflows (Proactive Engine)
start_plan, get_work, verify_and_advance, retry_with_improvement, create_workflow, advance_workflow, workflow_status
3. Task Management
create_task, update_task, list_tasks
4. Profiles & Workspaces
update_profile, workspace_write, workspace_read, workspace_list
5. Chat Branching & Managed Modes
fork_conversation, switch_branch, list_branches, set_conversation_mode, claim_manager, yield_floor, set_phase
6. Sub-channels
join_channel, leave_channel, list_channels
7. File Safety & Auditing
lock_file, unlock_file, log_violation
8. Shared Knowledge & Decision Tracking
kb_write, kb_read, kb_list, log_decision, get_decisions, get_compressed_history, get_briefing
9. Team Governance (Voting, Reviews, Feedback)
request_review, submit_review, call_vote, cast_vote, vote_status, request_push_approval, ack_push
10. Dependencies & Progress
declare_dependency, check_dependencies, update_progress, get_progress, get_reputation
Communication style
- Keep outbound messages to 2–3 short paragraphs.
- When work completes, acknowledge task IDs, files touched (by implementers), and blockers clearly.
Related skill
Developers implementing Neohive-assigned work should use neohive-developer-agent in this same .cursor/skills/ folder.