Runyard

Register or re-register an agent with the project and receive a summary of recent activity. Call this when an agent starts a session or wants a situational update. Returns tasks currently assigned to the agent, the 10 most recently updated memory entries, and the 20 most recent log entries across all agents for the project. After checking in, review your assigned tasks and move any you are actively working on to in_progress. When your coding/work is complete, move tasks to review (not done). Only move to done after review confirmation. Append knowledge (task_append_knowledge) as you work so the next agent has context. IMPORTANT: When your work is done, call agent_deregister to clean up. Do not leave stale agents registered. After checking in, check messages (agent_get_messages) for instructions from other agents. During work, use memory_set to store important discoveries (tag with the swarm ID if in a swarm), and use swarm_post_context to share findings with other swarm agents.

Remove an agent from the project when it is done working. Call this when an agent finishes all its tasks and is shutting down. This cleans up stale agent registrations and keeps the project tidy.

Get full details on a specific agent including system prompt and configuration.

Retrieve messages for a specific agent. Returns messages sent directly to this agent plus all broadcasts. Optionally filter by unread only.

Formally transfer a task from one agent to another with context. Updates the task assignee, appends handoff context as task knowledge, sends a message to the receiving agent, and logs the handoff.

List all registered agents in a project with their roles and last check-in times.

Post a log entry so other agents can see what is happening. Use this to broadcast progress, decisions, errors, or noteworthy events. Log entries are visible to all agents via agent_checkin.

Mark one or more messages as read by an agent.

Send a message to a specific agent or broadcast to all agents in the project. Use this for inter-agent coordination, requests, and status sharing. Use this proactively during work — not just at the end. Send updates when you find something important, hit a blocker, or need input from another agent.

Update an agent's configuration, system prompt, or role. Only provided fields are changed.

Dynamically load tool groups into this session. Call with no arguments or groups: ["list"] to see available groups and their tools. Call with groups: ["tasks", "memories", ...] to load specific groups. Reduces initial token overhead by only loading what you need.

Permanently remove a memory entry by key from a project.

Retrieve a specific memory entry by its key from a project.

List all memories in a project. Optionally filter by tags — only memories that have ALL specified tags are returned.

Full-text search across memory keys and values in a project. Returns all memories whose key or value contains the query string (case-insensitive).

Search memories using natural language. Uses vector similarity to find semantically related memories, even when exact keywords don't match. Requires the embedding engine to be available.

Store or update a key-value memory entry in a project. Use for FACTS, DECISIONS, and REFERENCE information — not for work items or plans. Good: architecture decisions, API conventions, environment notes. Bad: tasks, goals, initiatives, to-dos (use task_create instead). If the key already exists, its value is overwritten and updated_at is refreshed.

Create a new project container for organizing memories and tasks. Projects are the top-level grouping unit in AgentHub. Every memory and task belongs to exactly one project.

Get a project by slug with summary statistics including memory count, task counts by status, and registered agent count.

List all projects currently tracked by AgentHub.

Cancels a task by moving it to done with a cancellation marker in knowledge. Returns the updated task in protocol format.

Gets a single task by ID in the MCP Tasks protocol format. Returns protocol-level task status and, if completed, the task result with title, knowledge, and metadata.

Lists all tasks in the current project mapped to the MCP Tasks protocol format. Returns tasks with protocol-level status (working, input_required, completed, failed, cancelled). Supports cursor-based pagination.

Select or create the active project for this session. Call with no arguments to list available projects. Pass project_slug to select an existing project, or create_new to create and select a new one. Once set, all other tools use this project by default (no need to pass project_slug).

Show the current active project for this session. If no project is selected, tells the user to run runyard_init.

Execute a named skill (reusable multi-step workflow). Provide the skill name and all required inputs. The skill executor validates inputs, resolves template references between steps, executes each step sequentially against the database, and returns combined results. Use skill_list to discover available skills and their input requirements.

List all available reusable workflow skills. Skills are pre-built multi-step workflows that compose multiple Runyard tools into higher-level operations. Returns each skill's name, description, required inputs, and execution steps.

Register an agent to participate in a swarm with a designated role and optional budget limit. The agent will be added in "spawning" status. Swarm agents should follow proper task lifecycle: check in via agent_checkin, pick up assigned tasks, move them to in_progress when starting work, review when code is written, and done only after review passes. Do not skip kanban stages. Swarm agents should: check messages regularly, post discoveries to swarm_post_context, read swarm_get_context before starting work, and store important decisions in memory_set.

Advance the swarm to the next pipeline stage. If already at the last stage, the swarm is marked as completed. This tool checks for incomplete tasks — agents must move their tasks to done or review before the swarm can advance. Ensure all agents have followed the task lifecycle (in_progress → review → done) before calling this.

Create a new agent swarm with a goal, pipeline stages, and configuration. Swarms orchestrate multiple agents through structured pipeline stages (e.g. plan → execute → review → test). Returns the new swarm ID.

Read the shared context scratchpad for a swarm. Returns all entries posted by agents via swarm_post_context. Call this before starting work to see what other agents have discovered, decided, or flagged.

List all swarms in a project. Optionally filter by status.

Post a discovery, decision, or note to the swarm's shared scratchpad. All agents in the swarm can read this shared context via swarm_get_context. Use this to share findings, flag blockers, record decisions, or leave notes for other swarm agents.

Reassign a task from one agent to another within a swarm. If the from_agent's failure_count exceeds 3, their status is automatically set to "failed" (circuit breaker). Logs the reassignment in the agent_logs table.

Analyze swarm workload and return scaling recommendations. Checks task queues per status, agent utilization, and identifies bottlenecks. Returns recommendations like "spawn 2 more reviewers" or "reassign idle executor-01 to review". The master agent should call this periodically and act on the recommendations.

Start a pending swarm. Moves the swarm status to "planning" and sets the current stage to the first pipeline stage. Only works if the swarm is currently in "pending" status.

Get full swarm details including all registered agents and their states. Returns the swarm configuration, current pipeline stage, and agent breakdown.

Stop or cancel a running swarm. Sets the swarm status to "cancelled". Optionally provide a reason for the cancellation.

Update a swarm agent's status, budget usage, or failure count. Only the provided fields will be changed; others remain intact. Set status to "active" when starting work, and "completed" only after all your assigned tasks are in done or review status. Follow proper task lifecycle before marking yourself completed.

Add a knowledge entry to a task. As agents work on a task, they write back what they learn about the codebase, decisions made, and context discovered. This accumulated knowledge is available to any agent that picks up the task next.

Create a new task on the project kanban board. Use for ALL work items, plans, goals, initiatives, and actionable items. Tasks track work through a workflow: backlog → todo → in_progress → review → done. If you are planning something or tracking progress, this is the right tool (not memory_set). Create tasks in 'backlog' (default) or 'todo' for immediate work. Do NOT create tasks directly in 'in_progress', 'review', or 'done' — use task_move to advance them through the proper workflow. Move to 'in_progress' when you actually begin working on it — do not create tasks directly in 'done'. When work is complete, move to review with task_move (not directly to done). Only move to done after review passes. Only manage tasks assigned to you — never modify or delete other agents' tasks.

Permanently remove a task from the project kanban board.

Get full details on a specific task by its ID.

Retrieve all accumulated knowledge entries for a task. Returns the full knowledge base that agents have built up while working on this task.

List tasks in a project. Optionally filter by status, priority, and/or assignee. Returns all matching tasks sorted by creation time.

Move a task to a different status column on the kanban board. This is the primary way to advance work through the workflow (backlog → todo → in_progress → review → done). Follow the standard workflow: backlog → todo → in_progress → review → done. Move to in_progress when you START working, review when code is written, done only after review passes. Do not skip stages — never go directly from backlog/todo to done. Moving backwards (e.g., review → in_progress) is allowed for rework.

Update one or more fields on an existing task. Only the fields you provide will be changed; others are left intact.

Add a role to a team, binding an agent to a position in the workflow. In sequential workflows, the order determines task handoff sequence.

Advance a task to the next role in a sequential team workflow. Automatically reassigns the task to the next agent in the role order and moves it to in_progress. If the current agent is the last in the sequence, the task moves to done.

Create a team of agents with defined roles for structured collaboration. Teams enable automated task handoff between specialized agents (e.g. Architect → Builder → Reviewer).

Get full details on a team including all roles and workflow configuration.

List all teams in a project.

Remove an agent from a team by their agent_id.

Create an event-driven trigger that fires automatically when matching events occur. Triggers can send webhooks, create notifications, or log messages. Event types: task:changed, memory:changed, agent:checkin, agent:log, message:sent, message:read, team:changed, swarm:changed, project:updated. Conditions are ANDed — all must match for the trigger to fire. Empty conditions = always fire.

Delete a trigger by its ID. This permanently removes the trigger.

List all triggers configured for the current project.