How to use the Hosted Agents skill for remote coding agent infrastructure

What the hosted-agents skill does

A skill is a folder of instructions that an AI agent can load when a task matches its description. The hosted-agents skill is loaded when you are designing hosted or background agent infrastructure, such as sandboxed execution, remote coding environments, warm pools, session persistence, multiplayer collaboration, self-spawning agents, or Modal-style sandboxes.

Its core idea is that a hosted agent runs in a remote sandbox rather than on a local machine, and that session start speed should depend mainly on model time-to-first-token, with infrastructure setup finished before the user starts a session.

• Image registry: pre-build environment images with a cloned repository, dependencies and initial build steps, refreshed on a regular cadence
• Snapshots: save filesystem state after image build, after the agent finishes changes, and before sandbox exit
• Git identity: configure git explicitly in each sandbox and commit as the prompting user
• Warm pools: keep ready sandboxes for high-volume repositories and expire them as new images build
• Predictive warm-up, server-first agent design, per-session state isolation and real-time streaming
• Agent-spawned sub-sessions with start, status and continue primitives

Install the skill from a ZIP file

Download the ZIP for hosted-agents from the directory page, then unzip it. Put the resulting hosted-agents folder in one of two places depending on how you want to use it.

For a personal setup available across projects, place it in ~/.claude/skills/. For a single project, place it in that project's .claude/skills/ folder so it applies only there.

If you use the Claude apps, you can instead upload the skill under Settings → Capabilities → Skills.

• Personal: ~/.claude/skills/hosted-agents
• Project only: .claude/skills/hosted-agents inside your project

Trigger the skill with natural requests

You do not need a special command. Describe the design problem in plain language that matches the skill's scope, and the agent should load it. The skill's description lists the topics it covers, so mentioning them directly helps.

Requests that fit well mention infrastructure concerns, for example asking how to keep sandboxes warm, how to structure sessions for several clients, or how to let agents spawn sub-sessions.

• "Design a warm pool of sandboxes so coding sessions start quickly for our busiest repos."
• "How should I persist session state for a background coding agent that runs in a Modal sandbox?"
• "Plan a backend where Slack, a web app and a browser extension share one agent server."
• "Let my agent spawn sub-sessions for research across several repositories. What primitives do I need?"

Examples of what the skill produces

Because the skill is guidance rather than a tool, its output is design advice and structure. For instance, asked how to speed up sessions, the agent can work through predictive warm-up, building dependencies into the image, and letting reads start before git sync finishes while edits wait for sync.

The bundled references/infrastructure-patterns.md file includes a Modal integration pattern showing a sandbox class with snapshot restore. scripts/sandbox_manager.py is a Python sketch of sandbox lifecycle classes, labelled as pseudocode to adapt to a platform such as Modal or Fly.io.

What you need before using it

The included pseudocode script needs Python 3 and uses only standard library modules. Adapting the design to real infrastructure requires a hosted sandbox provider such as Modal or Fly.io. If you implement the git identity pattern, you need GitHub app installation tokens.

• Python 3 (for the pseudocode script only)
• A hosted sandbox provider such as Modal or Fly.io, if you build for real infrastructure
• GitHub app installation tokens, if you implement per-user git attribution

Safety notes: scripts and network

According to its metadata, this skill does not run scripts and does not use the network. The included sandbox_manager.py is pseudocode and will not run against a real provider unchanged, so treat it as a reference to read and adapt rather than something to execute as-is.

Review any provider credentials, tokens and sandbox permissions you add when you implement the design. The skill itself does not set these up for you.

• Runs scripts: no
• Uses network: no
• sandbox_manager.py is pseudocode and needs adapting before real use

Tips for getting the most from the skill

Use the skill for architecture questions about multi-user, hosted systems. For local use or one-off scripts, it adds complexity you do not need.

Ask adjacent questions to the skills that own them. The skill says research loops, sub-agent topology, tool design and file-backed session state belong to harness-engineering, multi-agent-patterns, tool-design and filesystem-context respectively.

• Name your scale: number of repos, expected concurrency and client types
• Ask for trade-offs, such as the cost of warm pools versus cold starts
• Keep sandbox infrastructure separate from clients, as the skill recommends