Sandcastle v0.43.0
Compare / Where Sandcastle Fits

Nine tools.
Three different jobs.

“AI orchestration” covers three categories that solve different problems, and most comparisons pretend otherwise. This page separates them, says plainly what Sandcastle is good at, and names the cases where you should pick something else. Every figure was verified from primary sources on 17 August 2026.

Verified ·· 2026-08-17
Sources ·· Official docs, GitHub, pricing pages
Method ·· No vendor claims taken on trust
01 / The categories

These tools are not substitutes for each other.

Durable execution engines

Temporal · Prefect · Airflow

Built to make work survive failure. A process dies mid-run and the workflow resumes where it stopped, sometimes mid-function. Decades of operational thinking, and the strongest reliability guarantees of anything on this page. AI features are recent additions on top.

Agent frameworks

LangGraph · LangChain · CrewAI

Built so a model can decide what happens next. You describe tools, state and control flow in code, and the model drives. Unmatched provider ecosystems and the richest primitives for agent loops, memory and tool calling.

Workflow orchestrators

Sandcastle · n8n · Dify

Built so a person decides what happens next, and the model does the parts models are good at. The workflow is a declared artefact — a file or a canvas — rather than a program. Sandcastle sits here.

The honest version: if a model needs to choose its own path through an open-ended problem, an agent framework fits better. If a workflow absolutely cannot lose state when a machine dies, a durable execution engine fits better. Sandcastle is for the large middle ground where the steps are known, the sequence is fixed, and what you want is for it to be readable, reviewable and cheap to run.

02 / The comparison

What actually differs.

Figures verified 2026-08-17. Star counts and prices move — check the linked source before you make a decision on them.

  Sandcastle LangGraph n8n Temporal Dify
Workflow is defined as YAML file Python / JS code Visual canvas Code, 7 SDKs Visual canvas (YAML DSL export)
License BSL 1.1 source-available MIT library · Elastic 2.0 server Sustainable Use fair-code MIT Apache 2.0 + conditions
OSI-approved open source No Library yes, server no No Yes Yes, with two conditions
Free self-hosted production use No until 2030 License key required Internal business use Yes Yes single-tenant
Workflows reviewable in git Yes plain files Yes it's code No PR review git on Business tier Yes it's code No git integration
Token & cost tracking built in Yes Via LangSmith paid Partial No Yes
Survives a crashed worker No run is marked failed Yes checkpointed Partial Yes resumes mid-function Partial
Human approval steps Yes gate / approval Yes interrupt() Yes Yes signals Yes
Local models first-class Yes ollama, vLLM, MLX Yes Ollama featured Yes n/a you write it Yes
Data residency enforced in code Yes EU No No No No
GitHub stars ~0.1k 39.9k 201.0k 22.4k 152.7k
Project age 5 months ~3 years ~6 years ~7 years ~2 years

The bottom two rows are the ones to take seriously. Sandcastle is a five-month-old project with one main author; every other tool here has years of production scar tissue and a community to answer your questions at 2am. That is a real difference and no feature column offsets it.

03 / Pick something else when

Cases where Sandcastle is the wrong answer.

The run must never lose state

Choose Temporal. Its event history replays a workflow back to the exact line it was executing, across process and machine failures. Sandcastle persists each step to the database and can replay a run from any step, but a worker that dies mid-step marks the run failed and a human decides what happens next. For payment flows or anything where a lost run means a lost customer, that gap matters.

The model should choose its own path

Choose LangGraph or CrewAI. Sandcastle has an agent step and a delegate step, but the graph is something you wrote down in advance. If the value of your system is that the model decides the sequence — branching into research it wasn't told about, calling tools you didn't anticipate — a framework built for that will fight you less.

The AI part sits inside a bigger data pipeline

Choose Airflow. If the LLM call happens between a warehouse extract and a load, splitting that across two orchestrators is a permanent tax on your team. Airflow's official AI provider now offers @task.llm and @task.agent with step-level caching, plus human-in-the-loop operators. It's young — version 0.7.0, first released April 2026 — but it lives where your data already is.

The people building workflows don't write code

Choose n8n or Dify. A YAML file is a developer artefact. If your ops or marketing team needs to build and change automations themselves, a canvas they can see is worth more than a file your engineers can review. n8n's integration catalogue is also an order of magnitude larger than Sandcastle's 64 connectors.

You need OSI-approved open source today

Choose Temporal, Airflow or Prefect. Sandcastle's source is published and auditable, but BSL 1.1 permits non-production use only until it converts to Apache 2.0 on 24 February 2030. If your policy requires a genuinely open licence for production right now, that is a hard blocker and no amount of goodwill changes it.

You are a Go, Java or .NET shop

Choose Temporal. Sandcastle is Python. Temporal has seven official SDKs, so your existing services keep their language.

04 / Where it does fit

Four cases where Sandcastle is the better tool.

The workflow belongs in code review

A Sandcastle workflow is a YAML file in your repository. It diffs, it branches, it gets reviewed in a pull request, it rolls back. n8n's own documentation states that its source control “doesn't support a pull request-style review and merge process” and that pulling “overwrites any changes you've made if they aren't stored in Git” — and git integration sits on a tier costing €667 per month. Dify has no git integration at all.

You need to know what a run cost

Cost and token accounting is in the engine and recorded per step, and a budget cap on the run warns at 80% and stops it at 100% — split evenly across parallel branches so a fan-out can't blow through it. Neither Temporal nor Airflow nor Prefect tracks LLM tokens or cost at all — that is code you write and maintain yourself. LangGraph routes it through LangSmith, which is a paid product with per-trace billing.

The data cannot leave the EU

Residency is enforced at the runtime layer, not promised in a subprocessor list, and local models are a first-class path rather than a fallback — ollama, vLLM and MLX endpoints run without any cloud key present. Combined with a tamper-evident audit chain and the Article-by-Article mapping on our EU AI Act page, that is a different starting position from a US SaaS with an EU region.

You want one file, not a framework

Twenty-five step types cover parallel fan-out, conditions, loops, gates, approvals, sub-workflows and deterministic steps that cost nothing to run. The deterministic ones matter more than they sound: most workflows are mostly plumbing, and plumbing shouldn't be billed per token.

05 / Licensing, without spin

Nobody in this category is simply “open source”.

It is the most misrepresented axis in the market, so here is the whole picture, including the part that is inconvenient for us.

Sandcastle · BSL 1.1

Source is public and auditable. Free for development, testing and personal projects; production use requires a commercial licence until the change date of 24 February 2030, when the whole thing becomes Apache 2.0. On the question “can I run this in production for free today”, this is the most restrictive licence on this page. What it buys you is a dated, contractual exit: one licence covering the entire repository, no separate enterprise directory, no branding requirement, and a guarantee it gets more permissive, never less.

LangGraph · MIT + Elastic 2.0

The library is MIT. The Agent Server you self-host is Elastic 2.0, needs a licence key, and reports usage to the vendor. RBAC, SSO and audit logs are enterprise features.

n8n · Sustainable Use

Not open source, and n8n says so plainly — they call it fair-code. Internal business use is permitted; offering it as a service is not. Enterprise-licensed files live inside the same repository.

Dify · Apache 2.0 + conditions

Genuinely permissive for single-tenant production. Two conditions: no multi-tenant service without a commercial licence, and the frontend keeps its branding. Contributors grant terms that permit relicensing.

Temporal · MIT

Unambiguous, OSI-approved, no strings. If licence purity is your deciding factor, this is the cleanest answer in the category.

Airflow · Apache 2.0

Apache 2.0 under Apache Software Foundation governance — vendor-neutral, which no commercially-backed project here can offer.

06 / Straight answers

Questions people actually ask.

Is Sandcastle open source?

The source is public and you can read, fork and modify it, but BSL 1.1 is not an OSI-approved open source licence — source-available is the accurate term. It converts to Apache 2.0 on 24 February 2030.

Can I replace LangGraph with Sandcastle?

Only if your graph is fixed. LangGraph exists so a model can pick its own route through a problem; Sandcastle runs a route you defined. Workflows with known steps port over well. Open-ended agent loops do not.

What happens if the server dies mid-run?

The run is marked failed and you decide what to do — replay it from any completed step, or fork it with different inputs. It is not resumed automatically. If you need automatic resumption across failures, Temporal is built for exactly that.

How does this compare to n8n on integrations?

It doesn't. Sandcastle ships 64 connectors covering 229 operations; n8n's catalogue is far larger. Sandcastle's argument is the workflow format, cost control and residency — not catalogue size.

Who is behind it and how big is the project?

It is a young project — five months old at the time of writing, with one main author and a small contributor base. If your procurement process weighs bus factor and community size, weigh them honestly here.

Do I have to use cloud models?

No. Local inference through ollama, vLLM and MLX is a first-class path in the runtime, and workflows run with no cloud API key present at all.

If the fit is right, it takes one command.

Try it against your own workflow.

$pip install sandcastle-ai Copy

Still unsure? Tell us what you're building and we'll say honestly whether this is the right tool — including when it isn't.