Monkey D LoopyAgent loops that know when to stop

Define a loop once. Prove it is bounded, deterministic, and resume-stable. Then compile it into a durable artifact that survives crashes without losing control of time, tokens, or cost.

Build your first loop

Give this to your agent

Browse recipes

Explore Gauntlet

🧱

Bounded by construction

Termination and caps are mandatory. Invalid or unreachable exit conditions are rejected before anything runs.

See the contract

🔁

Crash-resumable

Chained journals, deterministic replay, idempotent effects, durable sleep, and human breakpoints preserve control across restarts.

Understand the runtime

🧪

Proved before execution

Mocked-effect verification tests boundedness, determinism, and resume stability without shell, network, or model side effects.

Verify a spec

🧭

Agent-native

Use the CLI, the MCP server, or agent-readable documentation. Loopy checks grounding instead of trusting a model's claim that the work is done.

Integrate an agent

📦

Portable artifacts

Compile one LoopSpec into standalone, babysitter, Claude Code, Claude-native, and n8n targets—with an optional zero-install runtime bundle.

Choose a target

🛡️

Local-first operations

Run a secured local control center, schedule through one authority, index bounded artifacts, notify safely, and evolve through deterministic gates.

Operate loops

The load-bearing rule

Monkey D Loopy will not emit an unbounded loop. Every LoopSpec declares a termination signal and mandatory iteration, no-progress, token, dollar, and wall-clock controls. The validator also traces what feeds the exit predicate, so an agent cannot improve its safety score by merely relabeling its own opinion as external evidence.

BoundedThe loop has a reachable exit and hard limits even when the exit never arrives.

DurableEvery completed step can be reconstructed after a crash without repeating a completed effect.

GroundedReal evidence—not just the agent grading itself—determines whether the outcome is complete.

Quickstart

With Node 22 or newer, one command creates and completes a safe local loop—no model, API key, or external service required:

bash

npx --yes @loopyc/cli@latest quickstart

The quickstart scaffolds a LoopSpec, validates and scores it, executes one bounded iteration, reads the durable journal, and emits a zero-install standalone artifact. It leaves every result in ./loopy-quickstart/ so both a human and an agent can inspect what happened.

Continue with the two-minute first-loop guide, or install the CLI globally with npm i -g @loopyc/cli when you are ready to author real workflows.

Prefer a complete product workflow? Start from one of the verified recipes, which adds provenance, evidence requirements, safety boundaries, and adversarial fixtures on top of the same LoopSpec and runtime guarantees.

A useful mental model

text

goal or existing script
        ↓
   LoopSpec data  ─── validate ─── verify ─── score
        ↓
 compiled artifact
        ↓
 durable runtime ─── journal ─── inspect / resume / operate

Loopy is deliberately not a hosted workflow engine. It is a factory and runtime contract for one bounded agent loop. You keep the provider, coding agent, deployment environment, and artifacts; Loopy supplies the safety rails and durable execution semantics.

Documentation for humans and agents

This site is generated from the repository's canonical Markdown. Agents can consume the compact llms.txt index or the consolidated llms-full.txt context file. For the recommended authoring sequence and MCP integration, start with Using Loopy with agents.