Runtime – Sandboxed Coding Agents for Team Collaboration

TL;DR

TL;DR: Runtime spins up isolated OS-level sandboxes for coding agents so they can build and deploy freely without touching your actual machine or cloud.

Source and Accuracy Notes

⚠️ This section is MANDATORY. All links must be verified from actual source, not guessed.

• Project page: runtm.com ← MUST visit and verify
• Source repository: github.com/runtm-ai/runtm-coding-agent-runtime-control-plane ← MUST read README
• License: AGPL v3 (server) / Apache 2.0 (CLI) / MIT (templates) — verified via README badges
• HN launch thread: news.ycombinator.com/item?id=48225040

What Is Runtime?

Runtime is a YC P26-backed open-source project that gives coding agents — Claude Code, Cursor, Codex, Gemini CLI, and others — their own isolated OS-level sandboxes to work inside. Agents get full permissions to install dependencies, run tests, and deploy code, but everything stays sandboxed away from your actual infrastructure.

From the README:

Open-source sandboxes where coding agents build and deploy. Spin up isolated environments where Claude Code, Cursor, Codex, and other agents code and ship software. With live URLs, logs, and previews.

Setup Workflow

Step 1: Install

# Recommended
uv tool install runtm

# Alternative
pipx install runtm

# Or with pip
pip install runtm

Step 2: Start a sandbox session

# Start a sandbox (auto-installs deps on first run)
runtm session start

# Start with a web-app template
runtm session start --template web-app

# Use a different agent
runtm session start --agent codex

# List all sandboxes
runtm session list

Step 3: Attach your agent

# Reattach to a sandbox
runtm session attach sbx_abc123

Step 4: Deploy to a live URL

# Deploy from sandbox to live HTTPS URL
runtm session deploy

Deeper Analysis

How isolation works

Runtime uses bubblewrap on Linux and seatbelt on macOS for OS-level sandboxing. No containers needed — sandboxes start in under 100ms. Each sandbox is a fully isolated workspace with its own filesystem and process space.

Architecture

Your Agent  →  Sandbox  →  Runtm  →  Live HTTPS URL

1. Spin up a sandbox — your agent gets an isolated workspace
2. Agent builds — full permissions to code, install deps, run tests
3. Deploy — one command to a real URL with logs and previews
4. Iterate — agent can see logs, fix bugs, and redeploy

Infrastructure model

Machines spin down when idle and wake up on traffic. You get a live HTTPS endpoint on auto-stopping infrastructure.

Practical Evaluation Checklist

• [ ] uv tool install runtm completes without error
• [ ] runtm session start spawns a sandbox in under 100ms
• [ ] Claude Code or Cursor connects and runs inside the sandbox
• [ ] runtm session deploy produces a live HTTPS URL
• [ ] Sandbox persists files after runtm session stop
• [ ] runtm session destroy cleanly removes the sandbox and all files

Security Notes

OS-level sandboxing means agents have no direct path to your host filesystem or cloud credentials. The bubblewrap/seatbelt layer enforces strict syscall filtering. However: always review what your agent is doing before deploying, even in a sandbox — the sandbox protects your infrastructure, not the agent’s output quality.

FAQ

Q: How is this different from running agents in Docker? A: Docker images take seconds to spin up; Runtime sandboxes start in under 100ms. Runtime also has first-class agent CLI integration with embedded skill files for Claude Code, Cursor, and Codex.

Q: Does it work on macOS? A: Yes, Runtime uses seatbelt for macOS sandboxing and bubblewrap for Linux.

Q: What happens to my code if the sandbox spins down? A: Persistent workspaces preserve files across stop and resume. Destroying a sandbox deletes all files.

Q: Is there a hosted option? A: Yes, app.runtm.com offers hosted sandboxes. The core runtime is open source and self-hostable.

Conclusion

Runtime solves the trust problem for team-based coding agents. Instead of hoping an agent won’t rm -rf the wrong directory, you give it its own isolated world to play in. The workflow is clean — runtm session start, hand off to your agent, runtm session deploy — and the isolation is real OS-level enforcement, not just convention.

If your team is shipping coding agents to non-engineers or running agents on untrusted code, Runtime is worth evaluating. Self-host it on your own infrastructure or use the hosted option at app.runtm.com.