H
Hootan Moradi
Creator · Maintainer
AI engineer based in Amsterdam. Wrote the first line of FlowForge and still writes most of them. Works on orchestration, LLM primitives, and observability.
§01ABOUT
We build
the quiet infrastructure
under loud AI products.
§02 · MANIFESTO
We kept shipping the same glue code at every company. Retry loops. Idempotency keys. A cron job that kicks off a pipeline that kicks off a queue that eventually calls an LLM. A Slack alert at 3AM because step 4 of 9 ran twice and double-charged the customer.
FlowForge is what we wished existed: plain async functions with unreasonable durability guarantees, native LLM primitives, and no opinion about where you run it.
Open source is not a marketing checkbox. We've watched enough good tools disappear into acquisitions to know that your orchestrator should outlive whoever built it.
— Hootan & the team· maintainers, FlowForge
8
GitHub stars
early days · give it a ⭐
1
Maintainer
solo · built in the open
v0.4.1
Latest release
shipping every week
0
Dollars to start
MIT · self-host forever
§03TIMELINEThree years of shipping,
Three years of shipping,
one commit at a time.
- MAR 2024
First commit
A Friday afternoon and a frustrated rewrite of a Celery pipeline. The first
step.run()primitive fits in 80 lines. - AUG 2024
Public repo
Open-sourced under MIT. 400 stars in the first week. The first outside contributor opens a PR within 72 hours.
- JAN 2025
step.aishipsNative LLM primitive with retries, token tracking, and per-step cost attribution. Turns out durability is the killer feature for AI workflows.
- MAY 2025
Agent platform beta
Named agent identities, a task board with human-readable IDs, and a live SSE feed. AI team members become first-class.
- OCT 2025
v0.4.1 — latest release
Fair queueing, signed audit logs, tenant isolation. Four paying teams go live in the same week.
- NOW
Cloud, quietly
A managed option is in private beta. We're still focused on the open-source core — the Cloud is for teams who explicitly don't want to run infra.
§04MAINTAINEROne engineer.
One engineer.
Zero product managers.
FlowForge is built and maintained by a solo developer. If you file a bug, you will talk to the person who wrote the file.
§05OPERATING PRINCIPLESFive rules we argue over
Five rules we argue over
so you don't have to.
01
Open source, not open core
Every primitive you need for production is in the repo. No “enterprise tier” unlocking basic features. The hosted Cloud is convenience, not leverage.
02
Plain Python first
If you need a DSL, a YAML schema, or a special IDE plugin to understand a workflow, we've failed. async def and four primitives. That's the whole contract.
03
Durable by default
Every checkpoint, every retry, every crash recovery works without you configuring anything. “It might re-run” is not a valid footnote on a billing pipeline.
04
Observability is not optional
You can't debug what you can't see. Live SSE, full replay, per-step token and cost attribution — built in, not a paid add-on.
05
Ship on Fridays
Because we use FlowForge to deploy FlowForge. If we can't ship on a Friday, nobody can.
06
Answer the issue
Every GitHub issue gets a response from a human who wrote code, not a support bot. If that ever stops being true, we've gotten too big.