A control plane for your coding agent

Long agent jobs shouldn't die when your terminal does.

Smithers plugs into Claude Code and Codex and gives every multi-step run a durable spine: retries, approval gates, and full replay. You stay in the chat you already work in.

$ smithers claude installCodex too See it run

/workflows right in the chat Survives crashes & restarts Your models, your machine

claude code — ~/app

claude › /workflows

Smithers is driving. Pick a workflow:

▸ implement-and-review plan · build · review · verify

claude › run implement-and-review "add retry to deploy"

┌ run 7f3a durable · replayable

├ 1 plan ✓ done

├ 2 implement running

├ 3 review ○ waiting on your approval

└ 4 verify ○ queued

Close the terminal here and the run keeps going. Reopen and run 7f3a is still advancing, right where it was.

Why this exists

Your agent does great work, then forgets it happened.

Claude Code and Codex are brilliant at a single step. But real work has many steps, and the moment one of them slips, the whole thread unravels.

It dies when you close the tab

A sub-agent lives and dies with the terminal session. Shut the laptop, lose your connection, and the forty-minute job is simply gone.

One failed step starts over

Step three hits a flaky API. There is no retry and no checkpoint, so you rerun the whole thing from step one and pay for it twice.

No gate before the risky part

The agent runs the database migration without asking. There is no approval checkpoint and no replay to see what it did, only the aftermath.

How it works

One command, and the chat gets a durable spine.

Smithers installs as a plugin. It mirrors your workflows into a /workflows command and runs them on a durable engine that outlives the session.

Type `/workflows`

Your Smithers workflows show up right inside Claude Code or Codex. Pick one and the phase plan renders in the chat, derived live even from a run that already started.

It runs durably

Each step is checkpointed on the Smithers engine. Flaky steps retry on their own. Risky steps pause at an approval gate and wait for your yes.

It survives anything

Close the terminal, reboot, come back tomorrow. The run is still there. Reopen the live UI to watch it, replay it, or steer it.

Works where you work

Two agents, one control plane.

Same durable engine, native to each tool. Install the one you use, or both. Nothing new to learn, nothing to leave.

Claude Code

plugin + /workflows mirror

A native Claude Code plugin. Steer work to Smithers over ad-hoc sub-agents, get a live UI per workflow, and keep the phase plan in sync as the run advances.

$ smithers claude install

›_

Codex

plugin + live run UI

A Codex plugin that drives Smithers with a live UI for every workflow. The same durable runs, the same replay, from inside the Codex CLI.

$ smithers codex install

What the spine gives you

The parts your terminal never had.

◆ durable

Runs that outlive the session

Every step is persisted. The job keeps moving whether or not you are watching.

↺ retries

Flaky steps recover themselves

A transient failure retries instead of taking the whole run down with it.

⏸ gates

Approval before the risky move

Mark a step as human-gated and the run pauses for your call before it acts.

⤺ replay

Rewind and see what happened

Full time-travel over the run. Inspect any step, fork from it, or replay it.

◧ live UI

A window into the run

Open a live graph for any workflow, or watch it in the full-screen terminal monitor.

⚙ yours

Your models, your machine

Runs on the agent and models you already use. No new account, no data leaving.

Give your next long job a spine.

Install the plugin, type /workflows, and run something that can survive the night.

$ smithers claude install $ smithers codex install