Usher AI — open-source support agent with real tool access

Support that acts, not just answers

Role: Solo — architecture and build
Period: 2026 — in development
Status: In development

Most open-source helpdesks bolt an AI reply box onto a ticketing system. Usher AI starts from the agent instead — first-class MCP tool-calling means it can look up a real order or cancel a real subscription mid-conversation, rather than reciting a help article at you.

Highlights

Self-hostable — compose up plus an LLM key
MCP tool-calling as a first-class primitive
Postgres + pgvector as the only datastore

TypeScript
Mastra
Postgres + pgvector
MCP
Docker

The idea

Embed it on a website or inside a mobile app. It answers from the organisation's own docs through RAG, reaches into their live backend through their own MCP server, and escalates to a human the moment it is not confident.

Why it is different

The closest open-source options — Chatwoot, Tiledesk — are ticketing systems with AI attached. Usher AI is an agent with a ticketing escalation attached. That ordering is the product: the agent can change state in a real system instead of only describing how to.

Design principles

Simple by default, enterprise when needed — SSO, RBAC, voice and telephony are optional and off by default
Provider-agnostic across LLM, speech, telephony and identity
Docker-first from the first commit
Stateless application tier so scaling is purely additive

Stack

Node/TypeScript with Mastra as the agent framework
Postgres + pgvector for conversations, embeddings, RBAC and audit log
MCP client for live tool-calling
RAG ingestion for docs, FAQs and policies
Docker Compose with a bundled mock MCP server so tool-calling works on first run

Status

In development, MIT-licensed, built to be self-hosted. The repository is private for now.