NodeXLoop
Debug AI agents like you debug code. Build, run and govern agent loops on a
visual canvas inside VS Code: breakpoints, step-through, a full trace of every
call, a runtime guard, and loops that live as .loop.yaml next to the code they
drive. Runs against your own models (OpenAI, Anthropic, Google, OpenRouter,
Ollama or any OpenAI-compatible endpoint), with keys in the OS keychain.
Open it with NodeXLoop: Open Builder from the command palette, or from the
NodeXLoop icon in the activity bar. The full guide, covering every panel, is the
NodeXLoop User Guide (PDF).
Every panel
| Panel |
What it is for |
| Header |
‹ back, the ten tabs (the ones that don't fit go in More), project name and Save, ⋯ more actions (open, new, import/export JSON, export .loop.yaml, share link), undo/redo, the Ctrl+K command palette, Files, Integrations, Models, Architect, and Run loop, which is always on screen. On a narrow panel the secondary controls move into a drawer. |
| Compose |
Describe an outcome in a sentence; get a proposed loop from the template catalog, shape it in conversation, then Create or Create and run. Calls no model. A loop contract on the right lists participants, guardrails and checks. |
| Builder |
The canvas. Node palette on the left, the selected node's inspector (Settings, Memory, Logs) on the right, zoom and minimap bottom-left, status bar along the bottom. |
| Code |
The loop as its .loop.yaml, beside a pair that proposes diffs to accept or reject: iterations, pass score, pauses, renames, or your own hand edit. Accept goes through undo and rewrites the linked file. |
| Templates |
51 ready-made loops in nine groups, your saved templates, and starred ones pinned on top. |
| Runs |
Run history; each run as a Channel (the nodes talking, with threads for flagged steps and notes into the audit log), plus Output, Timeline and Logs, and an audit report per run. |
| Fleet |
A robot debugger that shows only the robots you deploy: wheeled AMR, quadruped, humanoid biped, tracked crawler or fixed arm, across concrete, carpet, wet floor, gravel, a ramp and stairs. Pause, step, slow motion, per-robot breakpoints (slip, near-fall, fall, stuck, unauthorized zone, joint limit, battery), speed limit, goals and patrol routes, fault injection, telemetry, JSONL trace export. |
| Orgs |
Sign in, join an organisation, submit loops for review, see released ones. |
| Autopilot |
Proposes simulated repairs for failed runs, inside the triggers, budgets, cooldowns and approval rules you set. Off until you turn it on. |
| Flow |
Fork a teammate's loop without touching the original; one attributed activity log per project. |
| Help |
Getting started, the model registry, keyboard shortcuts, and a link to the guide. |
| Dialogs |
Model registry, Files (run context), Integrations (Google, GitHub, Slack, signed webhooks to n8n/Zapier/Make, Control Room), Loop Architect (describe a loop, get a checked draft), the guard's block explanation. |
Getting around: the ‹ button, Esc (it closes whatever is open first),
Alt+← and the mouse back button all go back a tab. Every panel that collapses
at a narrow width has a button that brings it back as a drawer.
Only in the extension
- Live
.loop.yaml sync. Edit the canvas or the file; the other follows. Nodes
bind to the functions that implement them, with CodeLens and Problems-panel errors
when a binding breaks.
- Model calls from the extension host, so Anthropic and local Ollama work with
no CORS issues; workspace files as run context and output.
- Run, Search, Module and MCP Tool nodes. Run a command (off by default,
loopGraph.allowRunNodes), search through SearXNG, Brave or Tavily
(loopGraph.search.provider), install a governed Python package, or call a tool
on any server in your .mcp.json (loopGraph.mcp.enabled).
- Robot simulators from Hugging Face Spaces (NodeXLoop: Add Simulator from
Hugging Face Space…): pinned to a commit, hash-verified, built only when you
confirm, opened in a governed panel and shown on the Fleet as a live robot.
- SimReady assets (NodeXLoop: Get SimReady Asset for the Fleet…, or Fleet →
- Deploy): OpenUSD robots from the repos NVIDIA's SimReady Central lists. Choose
specs (metadata and physics layers, a few hundred KB), preview (plus
thumbnails) or full (the whole package, capped by
loopGraph.simReadyMaxMB,
default 256 MB). Public repos only, pinned, every file checked against the
published hash and the package's own BOM, nothing run, a receipt sealed.
- Kill switch (NodeXLoop: Set Agent Fleet Level…): L1 warn to L4 terminate
for agent runs and run nodes, from one governance file.
- NodeXLoop MCP server so Claude Code or Codex can build, check, simulate and
debug loops for you.
Commands
| Command |
|
| NodeXLoop: Open Builder |
The app, as a panel. |
| NodeXLoop: Reveal Binding in .loop.yaml |
Jump from code to the node bound to it. |
| NodeXLoop: Add Module Capability… |
A governed Python package as a node. |
| NodeXLoop: Add Simulator from Hugging Face Space… / Fleet Simulators… |
Import, build, open simulators. |
| NodeXLoop: Get SimReady Asset for the Fleet… |
Download a SimReady asset, with limits. |
| NodeXLoop: Set Simulator Escalation Level… / Set Agent Fleet Level (kill switch)… |
Governance. |
| NodeXLoop: Set Web Search API Key… / Set MCP Secret… |
Secrets, into the OS keychain. |
Settings
| Setting |
Default |
|
loopGraph.allowRunNodes |
false |
Let Run nodes execute commands (no shell, confined to the workspace, 5-minute timeout). |
loopGraph.mcp.enabled / loopGraph.mcp.configPath |
false / .mcp.json |
Let MCP Tool nodes start your MCP servers. |
loopGraph.search.provider / loopGraph.search.searxngUrl |
none / http://localhost:8080 |
Where Search nodes send queries. |
loopGraph.simReadyMaxMB |
256 |
Largest full SimReady download (ceiling 1024). |
loopGraph.pythonPath / loopGraph.nodePath |
auto |
Interpreters for modules and simulator builds. |
loopGraph.checkForUpdates |
true |
Daily check for a newer hand-installed .vsix. |
Development
npm install
npm run compile
Press F5 in VS Code to launch an Extension Development Host. The app's HTML in
media/ is generated by ../tools/build-guard.mjs; see the repository README.
License
Apache 2.0, the same terms as the rest of the Axiom repository, with LICENSE
included verbatim so the carve-out travels with the code.
Apache rather than MIT because MIT grants no patent rights at all. Apache 2.0
gives you an express, royalty-free patent licence for what ships here, which is
the point of using a permissively licensed tool built by a company that files
patents: you should not have to wonder.
Read the note at the end of LICENSE: the Two-Layer Evaluation Pattern
(Teacher-Student, Kid-Parent, Doctor-Patient) is not covered, and is
proprietary to Orivael Inc. Building a critic loop on this canvas is ordinary
use of the tool and is fine; reimplementing that pattern is what the note is
about. See TERMS_OF_SERVICE.md.
| |