| name | multi-engine |
| description | Route work to the right engine (Reactor, Sub-reactor, Agents) based on task characteristics |
| tags | ["routing","reactor","gemini","agents","efficiency"] |
| version | 1.0.0 |
Multi-Engine Routing
Three engines available. Route work to the right one based on task needs.
Engines
| Engine | Tool | Strengths | Cost | Trigger |
|---|
| Reactor | claude -p via bridge | Deep reasoning, complex code, architecture | Flat rate (time) | bridge.sh send |
| Sub-reactor | gemini-task | Web research, batch ops, file scanning | Free (1000 RPD) | gemini-task "prompt" |
| Agents | oc agent | Domain expertise, memory, tool access | Flat rate (Codex) | oc agent --agent <id> |
Decision matrix
Use Reactor when:
- Task needs host access (Docker, VPS, files outside container)
- Complex multi-step coding or debugging
- Source code analysis or modification
- Tasks that need Opus-level reasoning
Use Sub-reactor (Gemini) when:
- Web research (has Google Search built in)
- Bulk file scanning or auditing
- Tasks that are broad but shallow
- Reactor is in backoff/limp mode
- Need to preserve Reactor time for heavier work
Use Agents when:
- Task fits an existing agent's specialty
- Need domain knowledge (security, research, PM, GitHub)
- Task can be done inside the container
- Want Chartroom access and memory
Parallel execution
All three engines can run simultaneously:
oc agent --agent spec-research --message "Research X" --json &
gemini-task "Scan all SKILL.md files for missing versions" --dir /root/.openclaw &
MCP as shared protocol
Claude Code, Gemini CLI, and OpenClaw all support MCP servers. An MCP server built once works across all three engines.
OpenClaw native MCP (mcp-bridge plugin)
Agents can use external MCP servers directly — configured per-agent in openclaw.json:
{
"agents": {
"list": [{
"id": "spec-dev",
"mcp": {
"servers": [{
"name": "chartroom",
"command": "node",
"args": ["/path/to/chartroom-mcp.js"],
"type": "stdio"
}]
}
}]
}
}
- Supports:
stdio, sse, http transports
- Auto-namespaces tools:
serverName_toolName (prevents collisions)
- Per-agent provisioning = strict capability boundaries
- No MCP server mode (gateway can't expose itself as MCP server yet)
Priority MCP servers to build
- Chartroom (read/write/search) — replaces script-based
chart tool, gives agents native tool access
- Agent bus (post/read inter-engine messages)
- Config reader (safe read-only config access for agents)
Where MCP servers run
| Engine | MCP support | Config location |
|---|
| Claude Code | ~/.claude/settings.json mcpServers | Host |
| Gemini CLI | ~/.gemini/settings.json mcpServers | Host |
| OpenClaw agents | openclaw.json agents.list[].mcp.servers | Container |
| Reactor | Inherits Claude Code's MCP config | Host |
Rules
- Match engine to task — don't use Reactor for web research
- Free engines first — if Gemini or agents can handle it, save Reactor time
- Never block on one engine — use parallel execution when tasks are independent