Multi-agent support: run OpenClaw, OpenHands, Aider, etc. in sandboxes

[nthh]

Summary

Make remolt a generic sandboxed agent runtime — let users choose which AI coding tool to run, not just Claude Code.

User stories

• "I want to try OpenClaw without giving it access to my real machine"
• "I want to compare how Claude Code vs Aider vs OpenHands handle the same task"
• "I want a cloud-hosted OpenClaw I can access from anywhere"
• "I want the security of a sandbox for agents that need broad permissions"

Agents to support

• Claude Code (current default)
• OpenClaw (formerly Clawdbot/Moltbot) — open-source AI agent with 50+ integrations, 150k+ GitHub stars
• OpenHands (formerly OpenDevin) — autonomous AI software engineer
• Aider — AI pair programming in the terminal
• Continue — open-source AI code assistant
• Open Interpreter — natural language to code execution

Implementation

Agent selector in SetupForm

• Dropdown or card selector alongside existing repo URL field
• Each agent has: name, description, icon, required env vars (e.g., API key fields)
• Pre-selects Claude Code for backwards compatibility

Per-agent container images

Option A: Separate images per agent

• remolt-sandbox-claude, remolt-sandbox-openclaw, remolt-sandbox-aider, etc.
• Most isolated, each image optimized for its agent
• More images to build and maintain

Option B: Single image with agent install at boot

• Base sandbox image stays the same
• Agent installed/configured during inject_env phase
• Slower cold start, but simpler to maintain

Option C: Base image + agent layers (hybrid)

• Shared base image with common tools (git, node, python)
• Agent-specific setup scripts run at pod creation
• Ties into Project templates / starter environments #11 (project templates)

Server changes

• CreateSessionReq gets an agent field
• Server selects image or entrypoint based on agent choice
• Agent-specific env var validation (e.g., OpenClaw may need different API keys)

Security angle

This is a strong differentiator — OpenClaw in particular has drawn scrutiny for its broad permissions (email, calendar, messaging access). Running it in a remolt sandbox neutralizes those risks while still letting users experience the tool.

Considerations

• Each agent has different dependencies and setup requirements
• API key management: some agents use OpenAI, some Anthropic, some multiple providers
• Welcome message and entrypoint should be agent-aware
• Auth URL detection in useTerminal.ts may need agent-specific patterns
• Licensing: verify each agent's license permits this use