Tuff Lil Unit

A lil durable pipeline toolkit for AI coding agents.

Tuff's an ultra-simple implementation of the 'step function' pattern from Temporal and Inngest, built for AI coding agents like Claude Code and friends. Micro-scale, on your local machine.

It's just a little TypeScript and SQLite. Your agent could build this from scratch every time, but Tuff saves it the trouble — so it can focus on what you want to get done.

Tuff features

• Slot-based concurrency — new tasks start when a slot opens (vs. Claude Code's native Tasks = max 10 concurrent in batched waves, slowest holds flow).
• Three execution modes, mix freely — LLM API calls, Claude Code headless using your subscription (play at your own risk), or any async function.
• Concurrency per stage — slot limits per phase (fan out high for HTTP fetches, throttle back for LLM processing).
• Token budget — global and per-step limits (Claude Code subprocess kills mid-run).
• Progress and state is queryable — step results, token usage, and durations land in Tuff's local db. Talk to Claude about progress during execution.
• Domain storage — define your own data tables alongside Tuff's state tables.

An example pipeline

| Phase | What it does | Model | Calls | Concurrency | |-------|-------------|-------|-------|-------------| | Collect | Fetch 200 source URLs | none | 200 | 50 | | Extract | Pull structured data from each page | GPT-5-nano | 200 | 10 | | Classify | Score and categorise each result | GPT-5-mini | 200 | 5 | | Distil | Aggregate into final report | none | 1 | 1 |

~400 LLM calls, 200 HTTP fetches, 4 phases. One tuff() call.

Getting started

Install the skill:

/plugin marketplace add unfamiliar-city/marketplace
/reload-plugins
/plugin install tuff

Then use it:

Claude Code
Opus 4.6 · Claude API
~/Projects/myproject

 > /tuff build me a pipeline to make a million bucks.

Code example

The entire runtime is one function call:

// Pipeline 'my-pipeline', state persisted to ./state/tuff.db
await tuff('my-pipeline', { stateDir: './state' }, async (ctx) => {

  // Durable step — if this succeeded before, returns cached result instantly
  const data = await ctx.step('fetch', async () => fetchAll());

  // Fan out — one step per item, concurrent within slot limit
  // Each has a unique ID, so on resume only uncompleted ones re-execute
  const results = await Promise.all(
    data.map((item) => ctx.step(item.id, async () => process(item)))
  );

  // Fan in — summarise all results
  return ctx.step('summarise', async () => summarise(results));
});

Status

You either die a hero or you live long enough to see yourself reimplementing Kubernetes.

Alpha.

Not production-tested.

Claude Code CLI provider — a house of cards on top of undocumented Claude Code internals. Use if you're relaxed about robustness, approximate budget tracking is good enough, and you want to use your Claude subscription.

License

MIT