Munin

Munin

Your whole customer operation, behind one MCP endpoint.

Website · See it in action · Documentation · MCP Registry

CRM, conversations, outreach, journeys, CMS, knowledge base, and analytics on one Postgres schema — exposed as MCP tools your agents drive, not screens you click through. Headless the way a headless CMS is: the dashboard owns settings, auth, and the human half of the work — an inbox where a person takes over a conversation, and a review queue where they approve what the agent proposed — but there is no admin UI to click your way through the data itself. Every action runs through MCP tools, callable from any MCP-compatible client (Claude, Cursor, ChatGPT, custom runners) — same tools, same permissions, same audit log, whether a human or an agent is driving. Munin even ships its own: an in-process, per-org agent runner that answers live conversations and works the curation queue against an LLM provider you configure — so the platform runs out of the box, with external MCP clients optional.

The console — settings, auth, the inbox, and the review queue. It drives the same MCP tools your agents call.

The review queue — an agent drafts, files the evidence behind the draft, and waits; approving is the human's keystroke.

The embeddable chat widget — answering a live customer from the knowledge base, ready to hand off to a human and be picked back up by the agent.

Modules at a glance

These seven modules aren't separate products — they share one Postgres schema, one permission model, and one audit log. Together with integrations and platform plumbing they add up to 307 MCP tools and 75 skills. Watch how they tie together:

Core modules

Knowledge Base

• Markdown articles organized into spaces, each scoped to the audiences allowed to see it.
• Hybrid search that blends keyword matching with meaning-based results.
• Website import — crawl a public site and turn each page into an article in the background, automatically dropping articles when their source page disappears.
• Full version history with restore, plus a review queue for proposed edits.
• Agents revise a long article in place with targeted text replacements, instead of rewriting the whole body to change a sentence.

Conversations

• One inbox across email, chat widget, voice (Threll.ai / Vapi), and SMS (Twilio / MessageBird).
• Inbound *and* outbound — agents answer conversations and can place outbound calls.
• Images in and out on every channel: customers send screenshots, operators and the agent reply with them, and the agent can see what it was sent.
• Email that behaves like email — start by forwarding to a generated address, quoted history reconstructed from Gmail, Outlook and Apple Mail replies, auto-replies and bounces filed away instead of queued, junk senders remembered, undelivered messages surfaced and retried.
• Per-topic reply automation: the agent starts out drafting for approval on a topic, and is promoted to sending on its own once you trust it there.
• Assignable, organized by topic, searchable across every message, and filterable by status, origin, channel, topic and age.
• Built-in handoff to a human, with notifications to your own systems as conversations change.

CRM

• Contacts, companies, deals, activities, pipelines, and segments.
• AI-written summaries and suggested next actions, kept separate from what people edit by hand.
• Consent tracking — the lawful basis and source for each contact, required before they can be added to any outreach — with address deliverability tracked separately, so a bounced address never reads as a withdrawn consent.
• Automatic duplicate detection that proposes merges for review, plus bulk contact import.

CMS

• Content collections with structured fields, and entries you can publish in multiple languages, each locale with its own slug.
• Rich content blocks for article bodies — callouts, quotes, media, and more — addressable by key, so an agent can add, replace, remove or reorder a single block.
• Scheduled publishing and a media library that keeps your master image and serves sized derivatives.
• Full version history with restore, search, and cross-references between entries.
• A public content API serves your site or app, with draft preview links for custom frontends and engagement tracking built into every entry.

Outreach

• Propose-only outbound across email, SMS, and voice calls — campaigns, segments, and drafts for both first touches and replies.
• Broadcasts: one approved message to a segment, with the audience frozen at approval and a test send to check it beforehand.
• Org-wide send rules — quiet hours in the contact's timezone, a weekly cap, a cool-down after a conversation — apply to every marketing send, whether it comes from a campaign, a broadcast or a journey.
• Every proposal reports the exact address or number an approval would reach, and an approved send can be scheduled rather than fired immediately.
• Outcomes extracted from replies and from finished calls, written back onto the record.
• Recipients are drawn only from contacts who have recorded consent (see CRM).
• A person approves every message. The one exception is a journey step a person chose to let send on its own (see Journeys).
• Multi-step follow-up lives in Journeys; campaigns carry the first touch.

Journeys

• Versioned graphs of waits, branches, A/B splits and actions that move each person towards one goal, measured against a holdout group.
• People enter from a segment, from an event in your product (POST /v1/events) or a store order, or by hand; company journeys run once per company and message chosen people there.
• Steps send email, SMS, chat messages to signed-in widget visitors, and calls, update the CRM record, call your webhooks, or hand a person into another journey.
• Agent steps decide which path a person takes, fill in fields on the contact, or write a message for that one person from a brief and your knowledge base.
• Every message goes out through Outreach, so consent, unsubscribe, send rules and the review queue apply.
• Agents draft, validate and dry-run a journey; only a signed-in person publishes it, after a review that lists every email that would send without per-message approval. Fixed text can be set to send that way; agent-written steps earn it only after 30 reviewed drafts at 90% approved unedited; marketing texts and calls always wait for a person.
• Live journeys report goal conversion and split results, and a monitor pauses a step when its bounce, unsubscribe or failure rate crosses a threshold.

Analytics

• Captures page views and on-site searches across anything you want to measure, one row per view with read depth, reported per tracker.
• CMS pages are tracked automatically; for any other page, you add a small tracking snippet.
• Conversion funnels and per-visitor journeys — once someone is identified, their visits link to a CRM contact, including the anonymous ones from before. A signed email on identify means the website and the inbox resolve to one person.
• Breakdowns by traffic source, referrer, and country, plus "what to write next" signals (popular topics, engagement, and searches that came back empty).

Integrations

Three kinds of integration, three homes. Which one you want depends on what the other system *is*.

Messaging channels — where customers write to you

Email over IMAP/SMTP, or a generated forwarding address you point your existing mailbox at. The embeddable chat widget, themeable and dark-mode aware, speaking 23 languages. Voice through Threll.ai or Vapi, SMS through Twilio or MessageBird. Each one is a channel adapter behind the same conv_* tools.

Operator bridges — where your team already works

Slack mirrors each conversation into a thread: your team replies from the thread and the customer receives it, outreach proposals arrive with approve and dismiss buttons, and CMS publishes are announced with a link to the live article. The conversation, the approval, and the audit entry are the same ones the dashboard shows.

Data connectors — your customers' systems of record

Read live over the vendor's API and never copied into Munin, so there is nothing to sync and nothing to go stale.

Commerce and bookings have a self-service half: an end-user agent in the chat widget looks up *their own* order or booking, bound to the caller's identity server-side rather than to whatever email the model was told. Search Console is operator-facing and admin-only. Connections are stored encrypted, authorize by API key or OAuth redirect depending on the vendor, and are managed from the dashboard's Integrations page.

Have a system no vendor adapter covers? Point Munin at an MCP server you run and its tools become connector tools, with per-connection allow-lists over which of them agents may call.

Automation

Conversation loop

An in-process, per-org agent runner answers live conversations on every channel (chat widget, email, SMS, voice) against the LLM provider you configure — drafting and sending replies, and handing off to a human when needed.

Curator loop

The in-process agent runner also works a durable background job queue: scheduled KB curation, CRM hygiene, contact extraction, and outreach drafts, with retry and dead-letter handling.

Review

Everything an agent proposes and a person decides lands in one queue — KB revisions, outreach proposals, m