Swift · multi-model pipeline
let pipeline =
step → step → model
pipeline → result
agentflow kit
AFKit
Build the agent. A Swift SDK for pipelines that run natively on the device.
Jump to AFKit ↓
tools / engineering
AgentFlowSuite.
Run native agents on your device. Build them in Swift, give them a model made for your product, and ship them on Apple silicon with nothing sent to a server.
AFKit · AFStudio · optimized kernels
Two tools.One native stack.
One use case. A model of your own.
Describe
Select
Generate
Tune
Export
agentflow studio
AFStudio
Give it a model. From a described use case to a Swift package.
Jump to AFStudio ↓
agentflow kit
AFKit Pipelines,
cut to fit.
A Swift SDK for multi-step, multi-model agents that run on the device, alongside Apple's frameworks.
Shortest working example about six lines of Swift, tested on every release
A pipeline, end to end.
Step
Step
Model
Result
Chain steps and models into one object.
FEATURE
Pipelines as one object
[Chain models and steps, including models that are not language models.]
Draw the flow. Export to Swift.
Input
Model
Output
Connect
Export
FEATURE
A visual builder
[Draw a flow without code, then export it to Swift.]
Every step has a cost.
— ms
01
02
03
04
05
Illustrative · measurements pending
CHART
Where the time goes
[Latency for each step of a pipeline, on a named device.]
On device · AFKit
— mW
Cloud call
— mW
Same task. Same device.
Values pending
CHART
What a run costs
[Power and memory for one agent run, against the same job done by a cloud call.]
Build it in Swift. Keep it on the device.
agentflow studio
AFStudioFrom words toa Swift package.
Describe a use case and leave with a product-specific model as a Swift package.
A model made for the job.
01
Describe
Say what the feature should do, in plain words.
02
Select
Start from one of our base models.
03
Generate
AFStudio creates a dataset for the job.
04
Fine-tune
The base becomes a model for your product.
05
Export
A Swift package, with its budget sheet.
SWIFT PACKAGE
01
Describe the use case
What should the feature do?
Write it in plain words.
02
Generate examples
Inputs and expected outputs
A dataset for the job
03
Review the dataset
Examples matched to the use case
Ready for the fine-tuning step
AFStudio · workflow illustration
03 / Generate
A dataset built for the job.
01 Describe
02 Select
03 Generate
04 Fine-tune
05 Export
Describe → Select → Generate → Fine-tune → Export
underneath
Optimized kernels
The custom compute kernels every model and agent runs on, written for Apple silicon.
— mW
POWER
[task] on [device]
[One or two lines on what optimizing a kernel changes: the same model, less power.]
Read the researchTHE KERNELS
[Kernel 1]
[What it computes, and which models use it.]
[Neural Engine]
[Kernel 2]
[What it computes, and which models use it.]
[GPU]
[Kernel 3]
[What it computes, and which models use it.]
[Neural Engine]
[Kernel 4]
[What it computes, and which models use it.]
[CPU]
writing / engineering
How to make it fit on the device.
featured / engineering
[Featured post title, two lines at most]
[One or two sentences on what the piece shows and the number it ends on.]
Past what the toolscan reach?
A tighter budget, a harder device, or the design around the model: bring us one feature.
Talk to us