Avery NXR agents are live. Here's what's actually new.
· Avery NXR
We just shipped the Avery NXR agent platform. It's the biggest update to the product since launch — a complete local-first AI agent layer that builds both the agents and the apps they operate against.
Here's what's actually new, what it does, and the things that are most surprising about how it feels in practice.
What's new
An agent builder with 59 capabilities across 14 categories. Sub-agents, loops, conditions, knowledge bases, shell commands with guardrails, web access, file ops, integrations. Three creation modes: wizard, blank, or template. Visual drag-and-drop canvas with a YAML escape hatch.
7 production-ready agent templates pre-loaded on first launch. Anna (daily news), Sophia (meeting notes → action items), Marcus (resume screening), Priya (customer support triage), Carlos (sales pipeline digest), Yuki (competitor monitoring), Liam (server + endpoint health). Each one is a real working graph you can use immediately or fork.
63 connectors. 15 OAuth providers (Google, Microsoft, Slack, Discord, LinkedIn, GitHub, Linear, Notion, HubSpot, and more) and 48 API-key providers across 13 categories (email, SMS, web search, market data, commerce, databases, vector DBs, and more). One connection per service, reusable across every agent.
Multiple trigger types. IMAP inbox, scheduled, webhook, or agent-to-agent.
Consult Mode. Opt-in escalation to frontier models (Claude, GPT, Gemini) via your own BYOK keys, with anonymized and redacted payload preview before any data leaves your machine. Off by default. Per-task consent. Every escalation logged.
Excel / workbook bidirectional sync. Drop in your .xlsx, sheets become Prisma models, formulas become TypeScript, validation rules become Zod schemas. Bidirectional sync with row-level conflict resolution.
All of it running locally. The model is on your laptop. Nothing leaves unless you explicitly opt in.
What's most surprising in practice
A few things stand out from working with the agent layer for real:
The 90-second loop changes how you think about agents. When the cost of running an agent drops to electricity and the latency drops to local inference time, you stop being precious about which workflows are "worth" automating. Anything that repeats more than once a week becomes a candidate. The economics that used to gate "should we build an agent for this" disappear.
Templates compound faster than building from scratch. The 7 production templates are pre-loaded for a reason — they're the starting points most teams need. By the third or fourth agent you build, you're forking templates rather than starting blank. The templates become a delivery library, not just examples.
The audit ledger matters more than we expected. When every agent action is recorded — what triggered, what was considered, what was chosen, with what confidence — the conversation about "what is the AI doing?" becomes simple. We didn't fully appreciate this until we watched customers use it to answer compliance questions in 30 seconds that used to take hours.
Consult Mode hits less often than you'd think. The local model handles 95%+ of the operational work. The opt-in escalation to frontier models exists for the rare hard case. We're seeing single-digit percentages of tasks routed to Consult Mode in actual usage — which means the cloud LLM bill, when it exists at all, is small.
What this changes about how teams use AI
The agent layer turns Avery NXR from "a tool you use to scaffold a project" into "the operating system for your team's AI work."
You set up agents that run continuously, watch what happens, fork the ones that work, retire the ones that don't, and over time build a library of automated workflows that handle the operational layer of your business.
Most of our early users are reporting the same pattern: install Avery, run the 7 templates for a week, identify the 3-5 that actually deliver value for their workflow, fork those into customized versions, and add 1-2 new agents per week from there. After two months, they're running 15-25 agents handling the bulk of their daily operational work.
That's a different relationship with AI than "we have a ChatGPT Team subscription that some people use." It's structural automation, not occasional assistance.
What it costs
Free Desktop tier — $0 forever, no card required. Standard local models, 7 templates pre-loaded, standard connectors, full audit log.
Desktop Pro — $29/user/month. Premium local models, Consult Mode, cloud deploy targets (Vercel, Railway), premium templates and connectors, priority support.
Enterprise — custom pricing. On-prem deployment, SSO/SAML/OIDC, RBAC, compliance-ready templates, admin controls, dedicated support.
How to try it
Avery NXR is in private beta. Request access at avery.software/request-access. The 7 templates are pre-loaded on first launch — you can have your first agent running in under 5 minutes.
If you've been thinking about AI agents but weren't sure where to start, this is the lowest-friction version of "find out if it actually works for your workflow."
Mac · Windows · Linux. Local SLM via Ollama. Free Desktop tier means no commitment beyond installation.