BigBlueBam and the B-Suite

A work suite where AI agents work alongside people, with the same permissions.

Role
Founder and architect
Areas
AI, Tools, Apps
The BigBlueBam launchpad: a grid of tiles, one for each app in the suite.

Most work platforms are built for people talking to people, and an AI agent added later ends up poking at a UI that was never meant for it. We started BigBlueBam from the other end: one suite where a team and its AI agents plan projects, message each other, write docs and support customers in the same place, under the same permissions and audit trail.

Six cards: 19 products, 802 MCP tools, 148 automation events, 189 migrations, one front door, six link kinds.
Numbers counted directly from the source tree, not from marketing copy.

It started as a single project-management app and grew into the B-Suite: nineteen apps in one monorepo that share an organization, a login, a permission model and links between each other's records. A deal can become a project, a sticky note can become a task, and a won deal can become an invoice without re-keying anything. The full BigBlueBam manual (PDF) walks through every app.

Hub diagram with the entity links table at the centre and nine spokes showing which records link across Helpdesk, Bill, Bond, Bam, Book, Brief and Beacon.
These nine connections were backfilled from existing foreign keys, and agents add more through the entity-link tools, so a deal can be followed to its project and invoice.

The apps

  • Bam: project management with a Kanban board, sprints, and board, list, timeline, calendar and workload views.
  • Banter: team chat with voice and video calls, where AI agents post, schedule messages, subscribe to patterns and can join a call as a spoken participant.
  • Beacon: a knowledge base with hybrid semantic search, a knowledge graph, and per-article expiry that sends stale content back through verification.
  • Brief: real-time collaborative documents with versions and inline comments. Finished documents can graduate into Beacon.
  • Board: an infinite-canvas whiteboard with audio conferencing and a sticky-to-task pipeline.
  • Bolt: workflow automation, built either as WHEN/IF/THEN rules or on a visual node graph, triggered by events from every app.
  • Bond (CRM), Bearing (goals and OKRs), Bench (analytics), Bill (invoicing), Blank (forms), Blast (email campaigns) and Book (scheduling, including mixed human and agent availability).
  • Bay for media review with frame and region annotations and 3D model review, Bin for asset storage and structured-data editing, Blueprint for typed diagrams, Bureau for a virtual office with live presence, and Helpdesk for a customer support portal.
  • Blip: telemetry for your own apps. Post any JSON report and Blip discovers its fields, tails it live, redacts PII at the edge, and turns error spikes into Bolt events. Blip can also replace user and install IDs at ingest with codes that rotate daily, weekly or monthly, so teams can still count users and follow one install's issues without storing identifiers that link back to a person.
Bar chart of automation trigger events per source, from 16 for Bearing down to 1 each for Bench and Schedule, totalling 148.
A CI check rejects any published event missing from the catalog, so the list automations choose from always matches what the apps emit.

Agents as users

Our suite exposes more than 800 tools to AI agents through an MCP server, covering every app plus cross-app search, composite views of an account or project, and entity links. Agents work under the same authentication and audit log as people. Destructive actions go through confirmation tokens and a durable approval queue, and operators get per-agent kill switches and tool allowlists that fail closed.

Horizontal bar chart of MCP tools per product, from 109 for Bam down to 11 for Helpdesk, totalling 802.
Counted from the tool registrations in the MCP server source; Platform covers cross-app tools such as search, entity links, approvals and agent policy.
Six-step flow: authenticate, policy gate, permission check, confirm destructive actions with a short-lived token or queue them for human approval, then audit.
Three core tools (server info, who-am-I, heartbeat) stay callable even for a disabled agent so it can identify itself while an operator investigates.

How it's built

TypeScript throughout: React front ends, a Fastify API per app, PostgreSQL, Redis for real-time fan-out and queues, Qdrant for vector search, LiveKit for voice and video, and a Python voice agent. It ships with Docker Compose and a guided deploy script, so a team can run the whole stack on its own hardware.

We coordinate our own agent sessions with OrcVis.

Five stacked layers: clients, a single nginx edge, nineteen Fastify APIs plus an MCP server, worker and voice agent, a realtime layer, and a shared data tier of PostgreSQL, Redis, object storage and Qdrant.
The whole suite runs from one Docker Compose file, and every application container is stateless so it scales horizontally.