Provider/model lives on the node.
Every node stores a provider/model pair in its config — versioned with the canvas. The runtime resolves it the instant that node starts, so a change you make takes effect on the very next execution.
All features
models · per node
Model Hotswapping.
The provider and exact model bind to every node — swap either one mid‑run. Cheap models for fan‑out, frontier models for reasoning, across one chain.
Active
per-node provider + model
openai · anthropic · google · local
what it is
One model picker per node — not per project.
Most tools bind one model to the whole run. DISPATCH binds it to the node. Each node carries a provider/model pair in its config — change it in the inspector, or hotswap it mid-run, and the next node to execute picks up the new binding. No restart, no re-wiring, no code.
one chain · mixed models
bound per node
Investigate gemini-2.5-pro
Parallel llama-70b · local
Plan claude-opus-4-8
Create gpt-4o
Review o3
Five nodes · four providers · one Run.
try it
Swap a model on any node. Watch the run re-price.
Select a node, pick a provider and model, and the chain re-computes cost, latency, and provider mix — live.
model hotswap · live
chain running
the chain
tap a node to edit
create · inspector
provider + model
Models are real ids served by DISPATCH. Pricing is illustrative, per‑million tokens, to show how the roll-up reacts.
est. run cost$0.00
·
est. latency0.0s
·
providers0
provider mix
how it works
A binding on the node — resolved at execution.
Three small ideas make hotswapping safe and instant.
01 · binding
02 · adapters
One interface, every provider.
OpenAI, Anthropic, Google and any OpenAI‑compatible local endpoint sit behind one adapter. Nodes pass the same messages + tool schema; the adapter speaks each provider's dialect. Swapping is just pointing at a different adapter.
03 · live swap
Change it while it runs.
Already-running nodes finish on their current model; queued and downstream nodes pick up the new binding. No restart, no re-wire — the chain keeps flowing and just costs (or reasons) differently from here on.
resolve order · per node
node config → adapter → provider
node binding
Plan
anthropic/claude-opus-4-8
fallback: openai/gpt-4o
adapter
normalise messages + tools
stream · retries · token meter
provider call
claude-opus-4-8
streaming → downstream
why it matters
Right model, right node, right price.
cost
Spend where it counts.
Fan-out twenty agents on a cheap model; reserve frontier models for the two nodes that actually need to reason. One graph, a fraction of the bill.
Parallel · haiku
→
Plan · opus
resilience
No single-provider lock-in.
A provider is down or rate-limited? Hotswap the node to another and keep the run moving. Set a fallback once and the adapter handles the rest.
openai · 429
→
google · ok
privacy
Keep sensitive nodes local.
Route nodes that touch private data to a local model, while non-sensitive steps use the cloud. The mix is per node, so the boundary is exactly where you draw it.
Context · local
→
Doc · cloud
quality
A/B a node in seconds.
Suspect a node would do better on a stronger model? Swap it, re-run, compare the output and the cost side by side. Tuning the chain is one click per node.
Evaluate · gpt-4o
↔
Evaluate · o3
related
Pairs well with
Activeoffline
Local Models
Ollama, llama.cpp, LM Studio — fully local, alongside the cloud.
openIn developmentMODELS.md
Directive Model Control
One version-controlled file governs every model on the canvas.
openActiveusage
Analytics
Tokens and spend across every provider and local model.
open