Steps that survive anything
Completed work is memoized and checkpointed. Crashes, deploys, and retries resume exactly where they left off.
validatecached · 0ms
fraud-checkcached · 0ms
await-paymentrunning …
shippending
Workflows that forget nothing.
Retry anything.
Ship in an afternoon.
Event-driven orchestration with automatic retries, step memoization, and native LLM integration. Write plain async Python — FlowForge handles durability, fan-out, and the 3 AM pager.
- 01 Durable steps memoized across crashes & retries
- 02 Native OpenAI · Anthropic · any LiteLLM provider
- 03 Self-host on your infra — zero phone-home
run_id 8241_a2f7·attempt 2/3·elapsed 00:01:42p95 280ms
↓ event sourced
memoized ✓
§02PRIMITIVESPlain async functions.
Plain async functions.
Unreasonable guarantees.
Four step primitives. Memoization, retries, and state handled behind the scenes. Write Python; get durability.
workflows/order.py
# Memoized step — completed work returns instantly on retry async def process_order(ctx: Context): order = await step.run( "validate", validate_order, ctx.event.data, ) if order["total"] > 1000: await step.run("flag", flag_high_value, order) return order
§03LIFECYCLEFrom event to completion
From event to completion
in five acts.
- 01
Event arrives
An HTTP POST, a cron tick, a webhook, a queued message. Anything that fires in your stack can kick off a flow.
POST /events→order/created - 02
Function matched
FlowForge dispatches the event to every registered function whose trigger matches — by name, by filter, by schedule.
order/created✓ process_orderorder/created✓ notify_warehouseorder/created− send_receipt - 03
Run enqueued
A durable run record is written. Fair queueing ensures one noisy tenant can't starve the rest.
RUN 8241 - 04
Steps execute
Each
step.*call checkpoints to the store. Crash mid-flow? Redeploy? The run picks up exactly where it left off — no duplicate side-effects. - 05
Workflow completes
The function returns. The run is sealed, results stored durably, and every side-effect is logged with full replay data.
RUN 8241✓ COMPLETE
§04CAPABILITIESEight things
Eight things
you'd otherwise build yourself.
Cost attribution
Token spend per model, tenant, and workflow. No surprise bills.
$284.12
− 18% wk/wk
Live SSE stream
Every tool call, decision, and event as they happen. Not a dashboard — a tap.
0.2sfetch_user_context
1.1sAnalyzing sentiment…
1.8ssearch_knowledge_base
2.4sResponse drafted
Declarative flow control
Concurrency, rate limits, throttling, debouncing — one decorator.
@flowforge.function( concurrency=5, rate="60/min", throttle="10s", debounce="2s", )
Secure by default
Tenant isolation, scoped secrets, signed audit logs for every action.
Signing keyffk_live_•••• a2f7
Tenant isolationenforced
Audit logstreaming
SOC 2GDPRHIPAAISO 27001
AI agents as team members
Named identities with skills, models, and live status. Track them the way you track people.
A
Atlas claude-sonnet-4.5
researchcodereview
running
M
Maya gpt-4o
summarizetriage
idle
K
Kepler o1-mini
planverify
waiting
Human-in-the-loop
Approval queues for sensitive actions. Stay in control without stalling the pipeline.
Agent wants to issue refund of $420 to
ORD-9218Task board
Kanban with human-readable IDs. Assign to humans or agents.
Todo 2
FF-14
Sync customer data
Sync customer data
Doing 1
FF-12
Analyze churn
Analyze churn
Done 2
FF-09
Summarize Q3
Summarize Q3
§05LICENSINGFree forever. Your infra,
Free forever. Your infra,
your rules.
FlowForge is fully open-source. Clone it, run it, extend it — no license fees, no seat limits, no phone-home.
Self-hosted · MIT
Open Source
$0/ forever
Self-hosted. Full control. No limits.
- Unlimited workflows & runs
- All step primitives
- Event triggers & cron schedules
- Python & TypeScript SDKs
- Agent team platform
- Community support (GitHub)
Later this year
Cloud
TBDfair pricing, promise
A managed option for teams who don't want to run infra. Zero-ops deploys, shared skill registry, SSO.
- Hosted orchestrator (multi-region)
- Team workspaces & SSO
- Shared skill registry
- Managed secrets & audit
- Priority support
Q2Self-host GA
Q3Cloud private beta
Q4Cloud GA
§06QUESTIONS
Answers, before you ask.
How is this different from Temporal or Inngest?
FlowForge is Python-first with native LLM primitives baked in from day one — not bolted on.
step.ai() ships with retries, token tracking, and cost attribution at the step level, and the orchestrator understands agent identity. It's MIT-licensed, self-host-forever, and designed for teams shipping AI in production, not generic backend tasks.What does “durable” actually mean here?
Every
step.* call checkpoints its output to a durable store before continuing. If the process crashes, the machine is replaced, or you redeploy mid-flow, the next attempt resumes from the last completed step — without re-running side effects. It's the difference between “retry the job” and “retry the job and double-charge the customer.”Which model providers are supported?
Anything LiteLLM speaks: OpenAI, Anthropic, Google, Mistral, Bedrock, Azure, Ollama, and the long tail. Bring your own keys. Token usage and cost are attributed per step, per model, per tenant.
Can I run this on my own infra?
Yes — and it's the default. FlowForge is a single binary plus a Postgres-compatible store. Docker Compose, Kubernetes, a bare VM, a Raspberry Pi under your desk. No phone-home, no license server.
Is there a TypeScript SDK?
The Python SDK is the reference implementation. TypeScript is next and in active development — target parity with
step.run, step.ai, step.wait_for_event, and step.invoke.How do I migrate from a Celery / Bull / cron setup?
Incrementally. FlowForge functions coexist with your existing workers — point a few triggers at FlowForge, watch them run, move more over. Most teams migrate one pipeline at a time over a week or two.
— open source · self-host it anywhere
Clone the repo.
Ship your first durable
workflow in ten minutes.
$ pip install flowforge
$ flowforge init my-first-flow
$ flowforge dev
→ listening on :8288 · ready