huddle-mcp
A broker that queues agent requests (plans, questions, FYIs) and batches them by urgency into proposed meetings, leaving actual calendar booking to an attached calendar MCP.
README
huddle-mcp
An MCP server that lets your parallel coding agents schedule briefings on your real calendar instead of interrupting you. They queue their plans and questions; the queue batches them by urgency into time-blocks booked when you're free.
When you run a fleet of agents, the bottleneck is you. Three plans and two questions land at once, on their schedule, and the context-switching wrecks your decisions. huddle-mcp is the fix real employees already use: don't barge in — book a meeting.
auth-agent → request_meeting("JWT vs sessions", type=plan, urgency=blocker)
ui-agent → request_meeting("dark-mode default?", type=question)
db-agent → request_meeting("migration applied", type=fyi)
You: "book my agent briefings"
→ 🚨 Blocker · auth-agent: JWT vs sessions 11:05–11:15
→ 📋 Agent briefing · 2 items 14:00–14:20 (the question + the FYI)
You review at the booked time; the agents wait and unblock once you've decided.
The idea: a broker, not another calendar wrapper
Wrapping a calendar is a solved problem (Google ships an official Calendar MCP; there are mature OSS ones). The value here is the coordination layer in front of it:
- a shared queue any agent can drop a request into,
- an urgency-aware batching policy (blockers get booked ASAP; routine items bundle into the next briefing; FYIs ride along and reserve no time),
- a briefing renderer that aggregates everything into one agenda you skim.
So huddle-mcp is auth-free and provider-agnostic — it never touches your calendar. It emits the agenda + a desired time window, and the actual freebusy lookup + event creation is delegated to whatever calendar MCP you already have connected (Google, Outlook, …). Install = one line, zero new credentials.
Tools
| Tool | Who calls it | What it does |
|---|---|---|
request_meeting |
worker agent | Queue a plan/question/decision/fyi. Returns a ticketId. No calendar. |
check_response |
worker agent | Poll a ticketId; returns the user's decision once answered. |
cancel_request |
worker agent | Withdraw a ticket you resolved on your own. |
plan_meetings |
orchestrator | Batch the queue into proposed meetings + rendered briefings. |
confirm_meeting |
orchestrator | Record a booked calendar event; marks its tickets scheduled. |
get_briefing |
orchestrator | The rendered agenda + full tickets for a meeting. |
resolve |
orchestrator | Record the user's decision for one ticket → unblocks the worker. |
ingest_answers |
orchestrator | Parse the user's answers out of the event description and resolve every answered ticket at once. |
reopen |
orchestrator | Undo a decision and re-open the ticket (you changed your mind). |
discard_meeting |
orchestrator | Drop a meeting, re-queue its open tickets (stale proposals auto-expire after 24h). |
list_agenda |
orchestrator | What's queued / scheduled / answered. |
How it fits together
The calendar event is two-way: huddle writes the briefing into it, and the user answers right there.
worker agents ──request_meeting──▶ [ huddle-mcp queue ] (local JSON, file-locked)
(auth, ui, db…) │ batch by urgency + render briefing
▼
plan_meetings → proposed meetings (window + duration + briefing)
│
your orchestrator ──▶ your Calendar MCP: freebusy → create_event(description = briefing + answer block)
│
confirm_meeting(meetingId, eventId, start, end)
▼
┌── you answer in the event: "tkt_3f9a → use server sessions"
│
└─▶ orchestrator reads the event back ──▶ ingest_answers(description)
▲ (or resolve(ticketId, decision) in chat)
worker agents ──check_response(ticketId)─┘ poll, unblock, proceed
The briefing that lands in the event ends with a writable answer block:
— ✍️ Your answers —
Reply after each → then tell your agent "read my huddle answers".
tkt_3f9a2c1d → use server sessions
tkt_8b7e0a44 → ship dark mode as the default
When you say "read my huddle answers", the bundled huddle skill (in skills/huddle/) fetches the event via your calendar MCP, passes its description to ingest_answers, and every answered ticket resolves — the agents unblock. The same skill drives the booking side: "book my agent briefings" runs plan_meetings, finds free slots, and creates the events. huddle itself never touches the calendar — the skill bridges it.
Requirements
- Node.js ≥ 20
- A calendar MCP connected in the same client (e.g. Google Calendar) for the actual booking. huddle-mcp itself needs no account and no auth.
Install — Claude Code
{
"mcpServers": {
"huddle": {
"command": "npx",
"args": ["-y", "huddle-mcp"]
}
}
}
That's it — no env vars. (Prefer the bleeding edge? Use "github:marcoloco23/huddle-mcp" instead of "huddle-mcp".) Connect a Google Calendar MCP alongside it for booking, and (optionally) copy skills/huddle/ into ~/.claude/skills/ so "book my agent briefings" and "read my huddle answers" just work.
From source
git clone https://github.com/marcoloco23/huddle-mcp.git
cd huddle-mcp
pnpm install # builds via the prepare hook
node dist/index.js
For agent authors
Worker agents should queue and wait, never block the user inline:
request_meeting({ agent, title, body, type, urgency?, estimated_minutes? })
→ { ticketId }
# do other work, then:
check_response({ ticket_id }) → { answered, response } # proceed once answered
type:question(needs an answer) ·plan(review a plan) ·decision(pick an option) ·fyi(no action).urgency:blocker(book ASAP) ·normal/low(bundle into the next briefing). Defaultnormal.
Storage
The queue is a single JSON file shared by every agent process, at
~/.config/huddle-mcp/queue.json (honors XDG_CONFIG_HOME; override with
HUDDLE_HOME). Concurrent writes from parallel agents are guarded by an atomic
cross-process lock.
Developing & releasing
npm install # builds via the prepare hook
npm test # node:test — pure units + a stdio server integration suite
CI runs the suite on every push/PR. Releases are automated: bump the version in
package.json, then cut a GitHub Release — the publish workflow builds, tests,
and pushes to npm. It needs an npm automation token (which bypasses 2FA) saved
once as the NPM_TOKEN repo secret (Settings → Secrets and variables → Actions).
Roadmap
Where this is headed — autonomy (agents book themselves), closing the last mile (answered → agent actually resumes), notifications, smarter briefings, Outlook, and more — is in ROADMAP.md, ranked by how much it protects your attention.
License
MIT
Recommended Servers
playwright-mcp
A Model Context Protocol server that enables LLMs to interact with web pages through structured accessibility snapshots without requiring vision models or screenshots.
Magic Component Platform (MCP)
An AI-powered tool that generates modern UI components from natural language descriptions, integrating with popular IDEs to streamline UI development workflow.
Audiense Insights MCP Server
Enables interaction with Audiense Insights accounts via the Model Context Protocol, facilitating the extraction and analysis of marketing insights and audience data including demographics, behavior, and influencer engagement.
VeyraX MCP
Single MCP tool to connect all your favorite tools: Gmail, Calendar and 40 more.
graphlit-mcp-server
The Model Context Protocol (MCP) Server enables integration between MCP clients and the Graphlit service. Ingest anything from Slack to Gmail to podcast feeds, in addition to web crawling, into a Graphlit project - and then retrieve relevant contents from the MCP client.
Kagi MCP Server
An MCP server that integrates Kagi search capabilities with Claude AI, enabling Claude to perform real-time web searches when answering questions that require up-to-date information.
E2B
Using MCP to run code via e2b.
Neon Database
MCP server for interacting with Neon Management API and databases
Exa Search
A Model Context Protocol (MCP) server lets AI assistants like Claude use the Exa AI Search API for web searches. This setup allows AI models to get real-time web information in a safe and controlled way.
Qdrant Server
This repository is an example of how to create a MCP server for Qdrant, a vector search engine.