Novix Flow connects ClickHouse, Linear and your codebase to Claude. Today, in private beta, it turns error spikes into triaged Linear issues with a root-cause analysis and a proposed patch. Next: a fully autonomous agent that verifies every fix before you see it.
From error spike to a ready-to-review fix, inside Claude.
The private beta runs as an MCP server connected to Claude Desktop or Claude Code. Claude does the reasoning; Novix provides the tools, the context and the workflow.
- 01
Detect
Claude queries your ClickHouse error tables through the Novix MCP server and finds the spike.
Available in Beta - 02
Contextualize
Related Linear issues and the relevant source files are pulled into the conversation.
Available in Beta - 03
Reason
The triage-incident prompt guides Claude to a root-cause hypothesis and a proposed patch.
Available in Beta - 04
Write back
The Linear issue is created or updated with the root cause and the patch, ready for review.
Available in Beta
Built for the incident loop — today and next.
Every capability is labeled with its real status, so you always know what ships in the beta and what's on the way.
Native Model Context Protocol
Novix ships as an MCP server: tools for ClickHouse, Linear and your repo, plus a guided triage prompt. Claude does the reasoning inside the client you already use.
Hybrid model routing
Haiku triages every spike in milliseconds; only complex, multi-file cases escalate to Sonnet.
Prompt-cached repo context
Your repository's structure stays warm in Claude's prompt cache for fast, cheap repeated analysis.
Verified patches
Every patch runs your test suite in an isolated sandbox before a human reviews it.
Multi-agent response
Parallel investigators for logs, traces, deploys and dependencies, coordinated by an orchestrator.
Deploy correlation
Every spike linked to the exact commit or deploy that caused it, with a one-click rollback proposal.
One protocol between your telemetry and Claude.
Solid paths run in the beta today. Dashed paths are the autonomous agent we're building next.
The incident command center we're building.
A look at the planned hosted dashboard: live incidents, resolution trends and model routing in one place.
Live incidents
Demo dataMTTR · last 30 days
Model routing
- Claude Haiku82% · fast triage
- Claude Sonnet18% · multi-file synthesis
Design targets for Novix Flow.
Design targets, not measured results. We'll publish benchmark numbers as the beta matures.
Designed around Claude, end to end.
Fast, cheap triage on Claude Haiku for every spike. Deep, multi-file reasoning on Claude Sonnet only when it matters. Repository context kept warm with prompt caching, and every integration speaking MCP.
| Step | Model | Tokens / incident |
|---|---|---|
| Triage of each spike | Haiku | ~20k in / ~1k out |
| Root cause + patch | Sonnet | ~150k in (cached) / ~8k out |
Assumes 50 incidents/day per mid-size team (~1,500/month). Projection based on design assumptions, not measured usage.
Beta today. Autonomous tomorrow.
Private Beta
- MCP server for ClickHouse, Linear, repo
- Guided triage prompt
- Linear write-back
- Local-first
Autonomous Agent
- Haiku → Sonnet routing
- Prompt-cached repo context
- Sandbox-verified patches
- GitHub PR automation
Novix Cloud
- Incident Command Dashboard
- Multi-agent response
- Deploy correlation
- Enterprise & SOC 2 program
Join the private beta.
We onboard engineering teams running ClickHouse and Linear in small batches, with a 30-minute setup call and a direct line to the founder.
Questions? [email protected]