Avery NXR cookbook.

Paste an AgentCard URL, see the card preview, confirm. Two clicks.

1 min · 4 steps

Open the Runs tab + drill into the failing node. Every input + output is captured.

2 min · 4 steps

Four parts the wizard always wants: trigger, source, work, output. Name your providers explicitly.

2 min

Drop a User input node to pause the run and pop a form. Available only on manually-triggered agents.

2 min · 5 steps

Loopback / private-range URLs are rejected by default. Opt in explicitly when you need them.

1 min

Describe what you want in one paragraph, the wizard builds the graph.

3 min · 4 steps

The 'File Extract' node reads a document on disk (PDF, image, Word, scanned), runs layout-aware OCR, and optionally fills typed fields. For text already in the run, use 'Extract from text'.

3 min · 6 steps

The 'Extract from text' node reads any size of TEXT and pulls typed fields / records against your criteria. For files (PDF, image, Word) use 'File Extract' instead.

2 min · 6 steps

Drop files into the Knowledge tab; the agent's `kb.retrieve` nodes can vector-search them.

2 min · 3 steps

Rate a node or a whole run so the agent can learn what good looks like.

20s · 3 steps

Concrete before/after rewrites that show what the wizard wants to see.

3 min

When you describe an agent in plain English, the planner first tries to match your requirement to a pre-built template — match-and-fill is far more reliable than building from scratch. The fallback is the free-form planner; both end up at the same canvas.

2 min

The planner reviews + auto-fixes its own first-pass output before showing you the agent. The 'Verified ✓ in N passes' badge tells you what happened.

2 min

If your description uses Outlook for both reading AND sending, the wizard binds ONE Microsoft connection for everything email.

1 min

Re-fire a previous run with its original trigger payload — useful for email-triggered agents you don't want to wait on.

20s · 3 steps

Every Save creates a new version. The Settings tab's history lets you activate an older one.

1 min · 2 steps

Fire a one-off run regardless of the trigger schedule.

30s · 2 steps

Built an agent you'd reuse in a slightly different shape next time? Save it as a local template — the matcher will pick it up the next time you describe a similar agent in plain English.

2 min · 4 steps

Use natural-language schedules like "every weekday at 9am". Avery uses your local time zone by default.

2 min

Three controls in the page header: a search box that narrows the list as you type, a tiles/list toggle, and a sort picker.

30s · 4 steps

Switch an agent from manual to scheduled runs.

1 min · 3 steps

Save any agent to a single .nxragent.json file and re-create it on another machine in one click — no re-prompting the planner.

1 min · 4 steps

Drag nodes, connect ports, edit configs. The canvas is your agent's source code.

3 min

A freshly built agent shows a warmup card listing everything it needs configured. Required steps must be done before the first run; recommended ones are optional.

2 min · 4 steps

Only the message + configuration you author. Never secrets, OAuth, or project paths.

1 min

Agents are scheduled workflows that orchestrate the SLM, your connectors, and Avery's runtime to complete tasks autonomously.

1 min

Avoid "connect to X" — the wizard treats auth as setup, not a step. Avoid vague pronouns. Avoid stuffing with details.

2 min

Call a remote agent that speaks the open Agent2Agent (A2A) protocol from inside your own agent graph.

1 min

The wizard recognises common vendor names (Outlook, Gmail, Slack) and canonicalises them to OAuth slugs.

2 min

Drop an .xlsx, get a Next.js app with CRUD pages, charts, and bidirectional sync. ~90 seconds.

3 min · 6 steps

Scaffold a Next.js 15 app from a one-sentence description in about 60 seconds.

2 min · 4 steps

Concrete app descriptions land cleanly; vague ones leave placeholders the user has to fill in.

3 min

Concrete entity definitions + role lists matter more than length. Here's what the wizard actually uses.

4 min

The `xl.*` runtime gives you Excel's function surface in TypeScript. Formulas land in `lib/xl/<sheet>.ts`.

3 min

Edit on either side, push the diff back. The Database → Sync tab does both directions.

3 min

Entities mentioned in your description become Prisma models. Be specific.

3 min

Already have a Next.js project on disk? Avery can manage + extend it.

1 min · 3 steps

Six built-in themes; the planner picks a default per archetype but you can override.

1 min

When the .xlsx + the app both changed, the Sync tab's diff viewer shows you every conflicting row.

2 min · 4 steps

Open a project workspace and start the dev server with one click.

30s · 3 steps

SQLite for solo / local work. PostgreSQL for anything you'll deploy or share.

2 min

One entity per sheet, headers in row 1, consistent column types, no merged cells. Five minutes of cleanup = a much better app.

4 min

Each VBA module becomes a placeholder `.ts` file with the original code as a header comment + a per-feature translation guide.

4 min

Six spreadsheet-aware generators auto-inject. Here's what each one does + what shows up in the app.

3 min

Avery has 9 baked-in archetypes. Naming yours upfront tells the planner which generators to load.

2 min

Avoid implementation details ("use Tailwind", "add tRPC"). Avery picks the stack; you describe the product.

2 min

The /scaffold.clarify probe checks if your description / PRD has enough signal. Answer the questions to lock the plan; skip to fall through.

2 min

By default, scaffolded apps land in `~/nxr-projects/<slug>/`.

1 min

Six sections, ~500-1500 words. The wizard reads the whole thing — structure helps it pick the right archetype + entities + features.

5 min · 7 steps

Every generator emit, fix iteration, and migration lands in `.nxr/audit.json`. The Audit tab is your time machine.

2 min

Project-scoped automations run on a cron schedule against this app's runtime. Distinct from the top-level Agents page.

2 min · 3 steps

Upgrade, downgrade, cancel, swap payment methods, or pay a failed invoice. All billing flows go through Stripe Customer Portal.

2 min · 5 steps

The Code tab is a read-only file tree — quick inspection without leaving the workspace.

1 min

SMTP, Resend, SendGrid, Mailgun, Twilio, Tavily, Polygon, …

1 min · 3 steps

Microsoft (Outlook + Calendar), Google (Gmail + Calendar), Slack, GitHub, Linear — one consent flow each.

2 min · 3 steps

One Outlook connection covers every email-sending agent. No duplication.

1 min

Reusable, encrypted credentials your agents use to talk to external services.

1 min

Console runs `pnpm dev`, `pnpm build`, or any script in your `package.json` and streams output live.

1 min · 3 steps

Only when 3 local SLM iterations don't converge AND the request explicitly opts in.

2 min

The Database tab renders your Prisma schema as a diagram (or cards) so you can see every model + relation at a glance.

1 min · 3 steps

Your scaffolded app starts on SQLite. Switch to PostgreSQL once you're ready to deploy.

3 min · 3 steps

The Sync sub-tab keeps your scaffolded app in step with the source workbook — drift detection, push, and pull.

2 min · 3 steps

Ship to your own server. Avery rsyncs the build + runs your post-deploy script over SSH.

5 min · 3 steps

Same flow as Vercel — token, verify, deploy. Railway picks up your `railway.json` build config.

3 min · 3 steps

Verify your token, pick a project, and ship — the deploy stream runs in the Console tab.

3 min · 3 steps

The Library tab is the install surface; generators visible there auto-light-up in this project's Generators tab.

3 min · 3 steps

Pick a generator from the project-scoped list and run it again — useful after schema edits or when adding a new entity.

2 min · 3 steps

Generators are versioned plugins that emit Next.js code (CRUD, auth, dashboards, payments, …) against your Prisma schema.

1 min

Press Cmd+K, type, hit Enter. Every navigation + action is reachable by keyword.

1 min

Point Avery at a file or directory — chunks are embedded and stored in `kb_shared`.

2 min · 3 steps

Delete all chunks from a specific file or directory from the shared collection.

30 sec

Run an ad-hoc semantic query to verify what's indexed before wiring an agent.

1 min

The service is off by default — start it when you want to index documents or enable vector search.

2 min · 3 steps

A shared Qdrant vector database that lets every agent and generated app query the same indexed document set.

1 min

Pull a generator from the Available tab; it lands under `~/.nxr/plugins/`.

1 min · 3 steps

Generator plugins, agent templates, and connector kinds you can install on this machine.

1 min

Click 'Add MCP server' on Settings → AI → MCP Clients, pick the transport (stdio or HTTPS) + auth, and Avery creates the connector and opens the connection in one step.

2 min · 6 steps

Once connected, MCP tools appear in the agent builder palette under their server's group.

1 min

A stdio MCP server that connects fine in your terminal but fails in Avery is almost always a PATH / interpreter problem — Avery spawns it with a minimal environment.

2 min · 4 steps

Two AI sub-tabs, opposite directions. MCP Clients = Avery calls OUT to other tools. MCP Server = other apps call IN to Avery.

1 min

If you suspect the token has leaked, rotate it. The prior token is immediately invalidated.

2 min · 3 steps

An MCP server can do anything its underlying credentials let it do. Pick servers as carefully as you pick npm dependencies.

1 min

One click enables the server + mints a bearer token. Copy the token + the connection snippet into your MCP host.

3 min · 5 steps

Only what you opt in to. Off-by-default for every item; loopback-only; bearer-gated.

2 min

Model Context Protocol servers expose tool catalogs to your agents — GitHub, Linear, Postgres, Brave Search, anything that speaks MCP.

1 min

Expose Avery agents + capabilities to external MCP hosts (Claude Desktop, Cursor, …) over a localhost HTTP endpoint.

1 min

Claude Desktop's JSON config supports stdio-transport servers only. mcp-remote is a tiny shim that adapts HTTP into stdio.

1 min

Paste a .conf or fill in the fields, save, then toggle the tunnel on.

3 min · 5 steps

Common WireGuard bring-up failures + how to diagnose each.

2 min · 4 steps

A system-level VPN tunnel that lets your agents + generated apps reach databases or services behind a corporate WireGuard network.

1 min

WireGuard needs root to create a `utun` device and modify the system routing table. The prompt is the standard macOS admin dialog.

1 min

Browser callback timed out. Re-trigger the connect flow.

1 min · 3 steps

What each tier unlocks, when an upgrade pays off, and what stays the same across all plans.

3 min

Generators are re-runnable. Each run emits new files; user-zone comments preserve your edits.

2 min · 3 steps

Every generator emit is recorded in `.nxr/audit.json` with timestamps + the locked Plan id.

2 min

Database tab → Schema sub-view shows your Prisma models as connected boxes.

1 min

The Console isn't limited to dev/build/test. Type any package script.

1 min · 2 steps

Start `pnpm dev` from the Console tab and watch the output stream.

1 min · 3 steps

Each tab covers a distinct concern. Cmd+1..9 cycles them; j / k cycles forward/back.

2 min

Nine tabs that cover everything from running the dev server to inspecting the schema to deploying.

2 min

Scaffold runs through up to 5 fix-pipeline iterations on each generator. Activity Stream shows every step.

2 min

Avery scaffolds new apps under `$HOME/nxr-projects` by default. Move that.

30s · 3 steps

Avery NXR supports 7 cloud providers — Anthropic, OpenAI, Google, Google AI Studio, Azure OpenAI, AWS Bedrock, HuggingFace.

2 min · 5 steps

Add API credentials in one place — referenced by Build, Runtime, and Local Failover.

2 min · 3 steps

When Runtime is Local and an LLM call fails, prompt to retry with a cloud provider.

2 min · 3 steps

Choose which cloud provider + model drives agent + app generation.

1 min · 2 steps

Pick + install a Small Language Model — Avery's Runtime profile uses local by default.

5 min · 4 steps

Skip the local SLM. Route every `/llm.*` call to a cloud provider you bring your own credentials for.

3 min · 4 steps

Send reasoning to one model, completion to another, vision to a third. Per-role overrides let you mix and match.

1 min · 3 steps

Pick the skill folder, preview the translation, confirm. The new agent lands in /agents.

2 min · 5 steps

V1 refuses to import bundled Python or shell scripts. The translator replaces script-driven steps with deterministic action nodes where possible.

1 min

Anthropic's open Agent Skill format is a folder with a SKILL.md describing what the skill does. Avery imports it as an editable agent template.

1 min

Settings → AI → Test active. If the model crashed, restart Ollama / reinstall.

3 min · 4 steps

Avery NXR's first public release — local-first app and agent generation, with 7 agent templates, Excel ingestion, 63 connectors, three deploy targets, and an opt-in Consult Mode for frontier models.

3 min