Open source · durable runtime for AI agents

Agent workflows with time travel.

Zero config: describe the workflow in plain English and your coding agent builds it — no files written by hand. Then watch every step stream live, and rewind, fork, or replay any run from any point. Durable by default: kill the process, pick up where it left off. Any model, any harness.

Read the docs Docs for agents →Sign up for UI preview →

Run this in your project

$ bunx smithers-orchestrator initcopied ✓

Creates an editable hello.mdx workflow you can run immediately, then installs the smithers skill into your agents. From there, just describe any new workflow and your agent builds it — zero config.

ship-it · a7b93f2 running

you > Research, plan, then implement. Pause for my approval before you ship.
✓researchclaudepersisted
✓plancodexpersisted
↻implement (resumed)codexrunning
✓reviewpipersisted
⏸approval to shipyouwaiting
crash-safe · resumes from the last persisted step · rewind to any step

Model and harness agnostic

Claude CodeCodexCursorGeminiPi+ any AI SDK model

Works with the harnesses you already pay for: your Codex and Claude subscriptions, not just API keys. Point each task at whichever agent and model is best for the job, and swap the harness without rewriting the workflow.

runnable example workflows

ecosystem projects on the runtime

open source, no cloud required

Burns · Ralphinho · Cairo Coder · Agentix · Era · Local Isolated Ralph community projects on the runtime →

Why Smithers

The right way to build an agent
changes every six months.

Chains. ReAct. Crews. Swarms. Background agents. Subagent fan-out. Couple your infrastructure to any one of them and you’ve already rebuilt twice. One layer underneath never changes: durable orchestration. Smithers is that layer.

Claude, GPT, Gemini, Kimi

Volatile · weekly

ReAct, crew, swarm, background agents

Fluid · monthly

Prompting → loops → skills → recipes

Shifts · every 6 months

Durable steps, retries, state, approvals, observability

Stable · doesn’t change

Time travel · the killer feature

Every transition, attempt, and approval is a SQLite row, so you can watch a run live, rewind to any step, fork an alternate branch, and replay from a checkpoint. Full observability with Prometheus metrics and traces, no setup.

Every step is checkpointed the moment it finishes. A crash, a closed laptop, a flaky tool: all resume points.

Review loops, panels, debates, sagas. Named patterns are small TSX components on the substrate. Read them, fork them.

Score every run, then let Smithers optimize its own prompts (GEPA) against your eval suite. The system gets better, not just the answer.

Already fanning out subagents?

Your agent’s built-in orchestration is ephemeral.

Subagent fan-out inside Claude Code or Codex is the right tool for work that fits in one sitting. It also lives inside one session, one vendor, and one terminal. Smithers is where the fan-out goes when the work has to survive.

Built-in subagents

Work dies when the session ends or crashesOne vendor per sessionAn approval blocks the terminalA bad decision means starting overOrchestration is a prompt you retype

A Smithers run

Resumes from the last finished stepClaude, Codex, Gemini, and Pi in one workflowApprovals wait overnight, durablyRewind, fork, or replay any stepA workflow file you version, review, and rerun

vs. Claude Code Workflows →vs. Temporal →vs. LangGraph →

Just say this

Built-in recipes for everyday agent work.

Smithers ships a focused authoring pack plus a large archive of copyable examples. Describe the outcome and your coding agent can build the right workflow with create-workflow.

create-workflow create-skill docs-driven-development

“Migrate the dashboard to the new gateway APIs. Show me the plan before you start.”

→ runs mission: milestone plan · your approval · focused workers

“Add rate limiting. Don’t stop until tests pass and a reviewer signs off.”

→ runs implement: write · validate · review loop until approved

See what you can do →