Moss

Inspiration

Every team loses its story. A decision is made in a meeting, the follow-up lands in an email, the ticket goes to Jira, and a month later nobody remembers why. We wanted a place where that knowledge is captured once, connected, and reused, without anyone doing extra paperwork.

What it does

Moss is an enterprise knowledge platform run by a small grove of agents. Raven reads Gmail and Calendar, Firefly watches Slack, Fox handles Jira, Owl listens to meetings, and Tortoise keeps Confluence. Stag, the orchestrator, turns what they find into decisions, commitments and risks, then proposes the next step: a ticket, a Slack post, a calendar invite, or a draft email to HR, Finance or DevOps.

Nothing is sent without a manager's approval. Managers get a one- or two-sentence spoken update, a dashboard, and a Canvas where they describe a workflow in plain language and Stag draws it. Employees talk to the individual agents and see their own commitments.

How we built it

The backend is Python and FastAPI with Google ADK agents and Gemini as the main model. Memory has three layers: a cache for speed, a knowledge graph of typed, dated facts for connections across tools, and SQLite as the system of record. The frontend is React, TypeScript and Refine, with a clean notebook theme and a 3D enchanted grove built in three.js. ElevenLabs provides the voice and the creature sounds.

Challenges we ran into

Free-tier model limits forced us to build a model fallback chain with cooldowns and an offline mode. The model sometimes invented ticket numbers, so we added strict prompts and checks before anything is executed. Every connector also needed a mock twin so the whole flow could be tested without live accounts.

Accomplishments that we're proud of

One email can become a Jira ticket, a Slack message and a meeting invite, each approved with a single click, and the outcome is written back into memory. The Memory page shows side by side what each layer adds, so the three layers are something you can see, not just a claim.

What we learned

Human approval is what makes agents trustworthy at work. Memory matters more than the model: the same question gets a far better answer when the graph already knows who promised what.

What's next

Live meeting capture, real sign-in with roles from the company directory, and a hosted deployment for a pilot team.

Built With

• confluence
• d3.js
• elevenlabs
• fastapi
• gemini
• gmail-api
• google
• google-calendar-api
• jira
• python
• react
• react-flow
• refine
• slack-api
• sqlite
• three.js
• typescript
• vite