Your tools.
Our protocol.
Plug your CMMS, ERP, PLM, internal CRM, or anything that speaks JSON-RPC. NEXUPIRA is an MCP host — your servers become first-class tools the chat can call by name.
One spine · three shipped flows
Same chat. Three loops. All shipped.
Click a card. Each one cycles through a flow that's running in production today.
Automation drafts a Gmail and posts to Slack from a single sentence in chat. Vision sends a leaf photo through YOLO + a Vision LLM cross-check in twelve seconds. Knowledge answers from your own documents with citation chips, then asks the Memory Gatekeeper to remember what you taught it.
- Chat · Tool routing
- Workflow DAG
- RAG · Citations
- Memory Gatekeeper
- 8 connectors live
03.0 Use case stack · One spine, three loops
▶ Activate Edit canvas
Type a request. Nexupira parses it into a DAG workflow on the canvas. Hit activate — Gmail draft and Slack notification fire on schedule.
▶ Activate Edit canvas
What's the SLA penalty if a supplier misses the Q3 milestone?
03.1 Causal chain · Signal → trigger → action → audit
AUTOMATION
Chat-engine routes the request, the workflow-engine drafts a DAG, and on activation it lands a Gmail draft and posts to your Slack. Both connectors run in-process — no MCP subprocess hop.
VISION
Image goes to the BerryWise sidecar. YOLO returns bbox + disease list, optionally cross-validated by a vision LLM. The agent sees the merged result, not raw bytes.
KNOWLEDGE
Agentic RAG: the LLM decides when to search. Every answer carries citation chips that hover-preview the source. The Memory Gatekeeper runs three gates before persisting — generalisation, conflict, confidence.
03.2 Why different · Architecture as differentiation
Four properties competitors don't stack.
Most copilots have one of these. A platform has all four — that's when you stop building integrations and start shipping outcomes.
Your tools.
No RPC hop
between agents and tools.
Tool calls run in the same process as the chat engine. No microservice boundary, no network latency tax, no cold start. Survival on bad network days is a property of the architecture.
Every claim
links to its source.
Hover any 1 in the assistant's output to preview the source chunk. Click to open the document. Memory Gatekeeper validates before persistence — three gates, audit log on every store.
Pick the model
that fits your audit posture.
OpenAI, Anthropic, Groq, Gemini, Ollama, or any OpenAI-compatible Custom endpoint. Provider tag travels with every call into the audit log. Configuration-time choice, never the default.
03.3 Connectors · OAuth · per-user encrypted
Six tools out of the box. Any OpenAPI on top.
Connect Google Workspace, Slack, GitHub, Notion, Linear, and LINE on day one. Plug any other service through an MCP custom server. Tokens encrypted at rest, disconnect drops them atomically, every call lands in the audit log.
Google WorkspaceGmail · Drive · Docs · Sheets · Calendar
SlackRead channels · post messages · DMs · reactions
GitHubIssues · PRs · commits · CI · repo read-only
NotionDatabases · pages · comments · read & write
LinearIssues · projects · cycles · status updates
LINE MessagingOfficial account · broadcast · reply · webhook
Describe it. The substrate writes it.
Same chat for sourcing audits, release notes, supplier follow-ups, customer triage. Same DAG engine for sensors, schedules, webhooks. One system, infinite personas.