- OpenClaw drives Smithers through the native Smithers plugin and MCP tools.
- Smithers drives OpenClaw through
OpenClawAgentinside workflow tasks.
Install the native plugin
The fastest path ismcp add --agent openclaw. It installs the native Smithers
OpenClaw plugin into ~/.openclaw/extensions/smithers/, enables it in
~/.openclaw/openclaw.json, and registers the Smithers MCP server as a baseline
tool surface:
smithers_create_workflow, smithers_eval, and
smithers_optimize, plus a bundled smithers-orchestrate skill that nudges
OpenClaw toward reusable, eval-backed workflows instead of repeated one-off
turns.
Register MCP by hand
If you do not want the native plugin, OpenClaw stores MCP servers in~/.openclaw/openclaw.json under mcp.servers. Add Smithers as a stdio server:
list_workflows, run_workflow, watch_run,
resolve_approval, and the rest) become available to OpenClaw’s agent. Full
reference: MCP Server.
Optional generated skill files
OpenClaw discovers skills from~/.agents/skills/ (among other locations), which
is exactly where the cross-agent command installs them:
openclaw skills install <slug>);
the generated Smithers CLI skill files install directly into the skills
directory rather than through the hub.
Smithers runs OpenClaw too
In the other direction, Smithers can spawn OpenClaw as the worker for a workflow task throughOpenClawAgent:
bunx smithers-orchestrator init detects an authenticated OpenClaw runtime and can scaffold it in
.smithers/agents.ts even when Claude Code or Codex are not available. The
adapter shells out to openclaw agent --message <prompt> --json; pass session
or a per-call resumeSession to route work into a known OpenClaw conversation.
See also
- MCP Server: the full tool reference
- CLI Agents: using OpenClaw inside workflows
- Agent Support: the cross-agent commands