Agentic Automation RunTime
Agent authors.
Human approves.
Runtime executes.
Report proves.
aart is a runtime where your AI coding agent — Claude Code, Codex, Cursor — authors reusable, governed automation you approve once, and that then runs itself. Every run leaves structured evidence: per-step traces, artifacts, screenshots, console errors.
What it is
Agent-authored. Governed. Evidence-producing.
The lifecycle is: develop → validate locally → approve → run autonomously. The agent composes typed blocks and workflows; aart validates and executes them deterministically — no LLM involved at runtime. Approval is risk-based and lives at the artifact boundary, not per-step.
Typed blocks & workflows
The agent composes HTTP, browser, file, data, and CLI primitives into reusable workflow definitions with typed, validated arguments.
Verify reflex
Headless browser + page-render verification so the agent actually sees whether a page works, instead of guessing from code.
Governed CLI blocks
Read-only git, kubectl, docker, and gh commands run as a fixed binary and argv — no shell — with a full audit trail on every execution.
Structured evidence reports
Every run leaves per-step traces, pass/fail status, artifacts, screenshots, and console errors. Risk-based approval: low-risk runs free; riskier workflows need sign-off.
What's possible
Automation your agent can write today.
These are examples of what you can author, approve, and run right now with the current release.
Service health checks
HTTP reachability + response-time checks across endpoints, with a structured pass/fail report on every run.
Web & render verification
Navigate to any URL, assert visible text, capture screenshots, and surface console errors — confirms your pages actually render.
API & ops workflows
Compose multi-step HTTP + CLI operations — data fetching, transformation, and output — all in a single governed workflow.
Data pipelines
Parse JSON, YAML, and CSV; reshape and validate; write structured output to files — all without touching an LLM at runtime.
Read-only git / infra audits
Agent-authored git, kubectl, docker, or gh inspections that run with a full audit trail — governed read, no writes.
Scheduled & triggered runs
Kick off any approved workflow on a schedule or trigger — the runtime takes it from there, with evidence on every execution.
Install
Add aart to your agent in two steps.
Requires Node 20 or later. Browser features (verify reflex, web.read, screenshots) pull Chromium via Playwright on first use.
1. Install the CLI (optional but recommended)
npm install -g @team-monet/aartOr skip the global install and use npx @team-monet/aart anywhere.
2. Connect to your coding agent as an MCP server
Add this to your agent's MCP config (e.g. .claude/mcp.json, ~/.cursor/mcp.json, etc.):
{
"mcpServers": {
"aart": {
"command": "npx",
"args": ["-y", "@team-monet/aart", "mcp"]
}
}
}CLI commands
aart list List registered blocks and workflows
aart run Execute a workflow and get an evidence report
aart dashboard Open the local dashboard in your browser
Roadmap
What's here. What's coming.
aart ships as an open framework that rides existing infrastructure — MCP, CLI, npm, containers. No new console or GUI required.
- ✓Author → validate → test → approve loop, driven by your coding agent via MCP
- ✓Typed blocks & workflows: HTTP, browser, file, data, CLI primitives
- ✓Evidence reports: per-step traces, artifacts, screenshots, console errors
- ✓Verify reflex: headless browser render verification in one call
- ✓Governed CLI blocks: read-only git / kubectl / docker / gh with audit trail
- ✓Risk-based approval: low-risk runs free; riskier workflows need sign-off
- ✓Local dashboard for browsing blocks, runs, and reports
- ○Environments & promotion — dev → staging → prod with risk-based approval at the promotion boundary
- ○Portable deployment bundles +
aart serve— run autonomously on schedule, webhook, or event; retries, structured logs, OpenTelemetry - ○Sharing & distribution — publish blocks and workflows via npm or OCI; import community packs
- ○Broader agent-host support — first-class integration with additional coding agents beyond Claude Code and Codex
Get started
aart — Agentic Automation RunTime
Install @team-monet/aart, connect it to your coding agent as an MCP server, and tell the agent to author its first workflow.