scitex-agent-container — Declarative, On-Prem-First Agent Container Runtime

scitex-agent-container (sac) turns one spec.yaml into one long-lived, sandboxed, fleet-addressable Claude Code agent. Runs anywhere Apptainer runs — laptop, HPC login node, on-prem cluster, air-gapped server. Part of SciTeX.

Getting Started

• Installation
  • From PyPI
  • From Source
  • Requirements
• Quickstart
• Templates and Examples
  • Pattern Templates
  • Instantiating a Template
• Status and Hook Integration
  • Rich Status
  • Claude Code Hook Integration
  • Zero Coupling Design

Architecture & Reference

• How sac works
  • Launch flow
  • What each piece does
  • Restart / health
  • See also
• YAML Spec Reference (v3)
  • Quick links
  • Top-level shape
  • Field reference
  • Lifetime / session selection
  • kind: AgentProxy — HTTP forwarder agents
  • Examples
• Talking to a Running Agent
  • 1. A2A sidecar — POST /agents/<name>/turn
  • 2. CLI — sac agents send / sac agents tail
  • 3. sac listen — host-level HTTP control plane
  • Cross-host: sac --on <peer>
  • Forwarding to external A2A (kind: AgentProxy)
  • Picking a transport
  • See also
• Container Isolation in sac
  • Apptainer’s default leak paths (10 categories)
  • sac’s hardened defaults
  • Roadmap for sac’s isolation surface
  • One-line summary for papers / READMEs
  • See also
• Configuration and Runtime Directories
  • Layout
  • Configuration cascade
• Claude Code credential & settings files
  • Files
  • Fleet hosts
  • What NOT to emit
  • Safe metadata fields (whitelist)
• Apptainer Images
  • Builtin layers
  • Build
  • Sandbox / freeze workflow
  • Pinning a custom image

Ecosystem

• sac and orochi
  • The comms model — what sac knows
  • The corrected boundary
  • How orochi consumes sac
  • Standalone use
  • Live instance

API Reference

• scitex_agent_container API Reference
  • Config
  • Runtimes

Why sac

• Declarative agents. One spec.yaml per agent — the file is the agent (dir-as-SSoT). Reproducible, version-controlled, diff-reviewable.

• Rootless Apptainer isolation. Runs where Docker/cloud sandboxes cannot — HPC login nodes, on-prem, air-gapped. No root, no daemon, no Docker socket. Hardened by default via --containall.

• LLM-agnostic / on-prem-capable. Default: Anthropic OAuth. Alternative: any Anthropic-API-compatible backend (DeepSeek, MiMo / Xiaomi, a self-hosted LiteLLM / vLLM-with-Anthropic-shim gateway) via spec.claude.provider. Data, code, and inference can stay entirely on your network.

• Fleet ops out of the box. A2A push (POST /v1/turn), health / heartbeat, restart policies, multi-account credential rotation with auto-quota-watch, MCP + CLI + Python surface, cross-host orchestration via sac fleet.

• AGPL-3.0. Research-freedom license — infrastructure stays open, modifications stay shareable.

Quick Example

# Create and launch an agent from its YAML definition
sac agents start my-agent

# Fleet view / per-agent status
sac agents status
sac agents status my-agent --json

# Send a follow-up turn to a running agent
sac agents send my-agent "What is 2+2?"

# Tail the structured transcript
sac agents tail my-agent

# Stop / delete
sac agents stop my-agent
sac agents delete my-agent -y

Indices and tables

• Index

• Module Index

• Search Page