Skip to content

runloop.md

Latest commit

 

History

History
172 lines (118 loc) · 6.16 KB

runloop.md

File metadata and controls

172 lines (118 loc) · 6.16 KB
Raw

Runloop

"Your AI Agent Accelerator."
Website https://runloop.ai
Docs https://docs.runloop.ai
GitHub https://github.com/runloopai/api-client-python
Stars GitHub Stars
Classification agent-native
Category Code Execution Services
Compliance SOC 2 · HIPAA · GDPR (per marketing site)

Official Website

https://runloop.ai

Official Repo

https://github.com/runloopai/api-client-python — Python API client

https://github.com/runloopai/api-client-ts — TypeScript API client

https://github.com/runloopai/rl-cli — Runloop CLI (rli) including MCP server

How to Use (Agent Onboarding)

Interaction pattern: SDK / REST + CLI + MCP

  1. Register at platform.runloop.ai and create an API key (settings).
  2. Create isolated Devboxes (micro-VM sandboxes) and run commands, snapshots, and blueprints — Devboxes overview.
pip install runloop-api-client
# or: npm install @runloop/api-client  (see GitHub api-client-ts)
export RUNLOOP_API_KEY=...

CLI + MCP for assistants:

npm install -g @runloop/rl-cli
export RUNLOOP_API_KEY=...
rli mcp install   # Claude Desktop config
rli mcp start     # stdio MCP server

See Runloop CLI.

Agent Skills

Status: ⚠️ No official npx skills add … pack from Runloop yet.

npx clawhub@latest search runloop

MCP

Status: ✅ Available (via official CLI)

The Runloop CLI embeds a Model Context Protocol server so tools like Claude can list, create, and manage Devboxes — rli mcp start (stdio) or rli mcp start --http — Runloop CLI — MCP.

Detail Value
Docs https://docs.runloop.ai/docs/tools/rl-cli
Install helper rli mcp install for Claude Desktop
Devbox MCP injection Devbox create supports MCP gateway configurations (CLI reference)

What It Does

Runloop provides Devboxes: secure, isolated sandboxes for running AI-generated code and long-running agent workflows, with fast startup, optional suspend/resume, disk snapshots and branching for agent state, repo-aware environments, secrets/objects stores, and benchmark jobs to evaluate coding agents (e.g., SWE-Bench-style runs). The surface area is explicitly marketed for AI agents and AI software engineering workloads—not generic batch VMs only.

Why It Is Agent-Native

Criterion Evidence
Agent-first positioning Homepage: "Your AI Agent Accelerator" and "Launch AI agents on secure code sandboxes"; "Devboxes are code sandbox development environments… for… AI Agents" — runloop.ai
Agent-specific primitive Devbox lifecycle; snapshot/branch disk state for agent trajectories; benchmark jobs for agent evaluation; MCP wiring from CLI and devbox config
Autonomy-compatible control plane API-driven create/exec/suspend/shutdown; no human SSH session required for automation
M2M integration surface Python/TS API clients, REST, CLI, MCP via rli
Identity / delegation Per-devbox IDs; API keys; secrets injection; audit-friendly logging (per product docs)

Primary Primitives

Primitive Description
Devbox Isolated execution environment for agent code (micro-VM + container isolation, per product)
Snapshot Capture and branch disk state for reproducible agent runs
Blueprint Reusable image/template for agent environments
Benchmark job Orchestrated multi-agent evaluation against public or custom scenarios
Object / secret store Agent artifacts and credentials without embedding secrets in prompts
Suspend / resume Cost-aware pausing of bursty agent workflows

Autonomy Model

Agent obtains RUNLOOP_API_KEY (or delegated token)
    ↓
Creates Devbox from blueprint or snapshot via API/CLI
    ↓
Runs commands, tests, or full agent loop inside sandbox
    ↓
Snapshots state for regression tests or parallel branches
    ↓
Shuts down, suspends, or resumes without human provisioning

Identity and Delegation Model

  • API keys — tenant-scoped access to Runloop API
  • Devbox IDs — isolate sessions and attribute logs
  • Secrets — injected as env vars without surfacing values to the LLM (pattern documented in CLI)
  • Network policies — optional egress and MCP gateway controls (CLI reference)

Protocol Surface

Interface Detail
Python SDK pip install runloop-api-client — api-client-python
TypeScript SDK @runloop/api-client — api-client-ts
REST API Platform API per docs.runloop.ai
CLI npm install -g @runloop/rl-cli → rli
MCP rli mcp start / rli mcp install

Human-in-the-Loop Support

Optional review of agent outputs or benchmark results. Core sandbox execution does not require a human in the loop.

Why Generic Alternatives Do Not Qualify

Alternative Why It Fails
Raw cloud VMs No agent-first lifecycle, snapshot branching for trajectories, or built-in agent benchmark orchestration
CI runners Job-centric; not optimized for interactive agent loops, suspend/resume, or MCP-native assistant control
Local Docker Host risk and ops burden; lacks managed agent evaluation and fleet scale patterns

Use Cases

  • Coding agents — run, test, and iterate generated code in isolated Devboxes
  • Agent CI / regression — snapshot state and re-run after model or prompt changes
  • Benchmarking — compare agent implementations on public suites at scale
  • Long-horizon tasks — suspend/resume bursty workloads to control cost