1
Run one command
From inside your project:In a terminal,
init asks exactly one question: which coding agent do you prefer? It shows what it detected (Claude Code, Codex, Pi, …), you pick one, and it does the rest - scaffolds .smithers/ with the focused authoring pack (create-workflow, create-skill, and docs-driven-development), installs dependencies, and installs that agent’s Smithers plugin (or skill). Former recipes are archived under examples/init-pack/. Piped or non-interactive callers (CI, agents, --yes) skip the question and install the defaults; --no-tutorial skips the guided session.The star of the scaffold is create-workflow - describe the workflow you want, and it writes the workflow plus a companion agent skill document. Re-running init later only offers to update files that drifted.2
Make sure your agent knows Smithers
Your agent drives Smithers, so it needs the Smithers context. Two ways to get it there:
- Have your agent read the docs. Ask it to run
bunx smithers-orchestrator docs(concise index) orbunx smithers-orchestrator docs-full, or point it athttps://smithers.sh/llms-full.txt. Quick, no install. - Install the plugin or skill for your harness so the context is always loaded. Each page below has the exact plugin/skill install:
init already installs the best integration for your preferred agent - the Claude Code plugin, the Hermes/OpenClaw native plugins, or the curated skill for Codex, Pi, OpenCode, Kimi, Amp, and Antigravity. Use the page for your harness above to wire additional agents or add MCP.3
Ask your agent to create a workflow
You don’t run Smithers yourself - your coding agent does. Ask it, in plain language:
“Use create-workflow to build a workflow that checks my API and keeps iterating until tests pass.”Your agent runs
create-workflow, a durable run completes, and it reports the generated workflow and companion skill path back to you.When you run it yourself in a terminal, a cheap, fast model narrates the result: a short recap in the terminal plus a self-contained HTML summary that opens in your browser (written under .smithers/reports/). Pass --no-report, or set SMITHERS_NO_REPORT=1, to skip it.This runs the coding agent
init detected (Claude Code, Codex, …), so have one installed and signed in. From here on, your agent drives Smithers for you - you describe outcomes, it picks and runs the workflow.4
Ask your agent to open the run UI
To watch a run live in your browser, ask your agent:
“Open the Smithers UI for that run.”Your agent runs
bunx smithers-orchestrator ui <RUN_ID>, which opens the run’s live status card - stages, output, and any approval gates - in your browser. Keep it open while you hand off more work.Read next
Your agent ran a workflow. From here, pick by what you need next.Talk to your agent
How-to: phrase requests so your agent picks the right workflow every time.
Watch and steer
How-to: follow a live run, clear approvals, and redirect without touching code.
What you can do
Reference: the full catalog of outcomes you can hand off, and what to say.
The few concepts you need
Explanation: the handful of ideas behind durable runs, gates, and loops.