Nucleus + Dispatch

The same workflows, in the cloud or on your box.

Both run workflows. Nucleus runs them in the cloud, so you sign in and build, with nothing to install. Dispatch runs them on your own hardware, with local models, your own MCP servers and the systems on your own network. Every tool call either one makes goes through Nucleus's governance, so the choice is about where the work runs, not about the rules it follows.

Both orchestrate. The difference is where.

Nucleus and Dispatch both turn a sequence of tool calls into a workflow that runs without anyone watching. Nucleus does it in the cloud, which is the quickest way to start. Dispatch does it on hardware you own, which allows things a cloud service can't offer.

In the cloud

Nucleus

“Sign in and build.”

Orchestrates: Workflows built in the browser and run in the cloud
Tools: Everything in the catalog: the company's systems, already connected and credentialed
Models: Hosted models, metered per use
Governs: Every tool call from every client, Dispatch included: default-deny, rate limits, budgets, ledger and audit
Best for: Getting started. There is nothing to install and nothing to run.

On your own box

Dispatch

“Own the hardware, and the freedom that comes with it.”

Orchestrates: The same kind of workflows, plus agents, schedules, triggers and memory
Tools: Nucleus's catalog, plus local MCP servers you run yourself
Reaches: Systems that only your own network can reach, such as file shares, internal databases and devices
Models: Local models on your own GPU, with no per-token bill
Best for: High-frequency work, private data, and anything that has to stay on site

Start in Nucleus. When a workflow needs a local model, a local MCP server or a system on your own network, move it to Dispatch. It still calls the same Nucleus tools, under the same names, rules and audit trail, because Dispatch reaches the company's systems through Nucleus rather than around it.

How a request travels

This is the path when the work runs on a Dispatch box. It serves a person typing in Claude, a tap in the iPhone app, a Slack command and a 3 a.m. schedule. A Nucleus workflow follows the same path without the box: it runs in the cloud and goes straight to the governance layer.

Where you areSurfaces

ClaudeDispatch iPhone appSlack slash commandWebhookSchedule

describe it · start it · approve it

Decides whenDispatch

Workflow DAGSchedules & triggersApproval gatesLocal modelsRun history

MCP tool call · as the workflow's owner

Decides whetherNucleus

IdentityDefault-denyRate limitsCredits & budgetLedger & audit

credentialed call · on that person's behalf

Does the workSystems

Translation platformFlatten cacheCRM & billingTicketsChat

Spend tokens only where judgment is needed

This is where running on your own box pays off. Most recurring operational work needs no judgment, just the same fetch, filter, compare and write every time. A workflow decides at each step how much intelligence that step needs, and most steps need none. For the steps that do, Dispatch uses a local model with no per-token bill.

per step 0 tokens

Deterministic tool step

Fetching, filtering, joining, diffing, counting and writing, done by typed tools. These steps never call a model, and the same inputs give the same output every time.

Use for everything that has a right answer

per step electricity

Local model step

A model on the box's own hardware handles the step that needs judgment, such as summarizing, classifying or drafting. No per-token bill, so running hourly costs no more per run than running weekly.

Use for the one step that reads or writes prose

per token metered

Frontier model

The strongest models, used mostly at design time: describing a workflow in Claude, debugging a run, or changing a step. You pay for that intelligence once rather than on every scheduled run.

Use for designing the automation, not repeating it

Once

Design it with Claude

Describe the job. A frontier model builds the workflow over MCP, runs it against real data, and you fix one step at a time until the output is right.

Every hour, after that

Run it with no model at all

The finished workflow is just tools in a fixed order. It runs on a schedule, calling Nucleus for every system it touches and adding nothing to anyone's token bill.

0model calls

across the cache-freshness workflows for one customer, every run

28scheduled runs / day

an hourly requeue plus a sweep of six locale sites every six hours

~170tool calls / day

about 125 of them go through Nucleus, each authorized, metered and recorded. The rest are local data steps.

$0tokens / run

the running cost is the electricity for the box

Nucleus makes the spend visible, and Dispatch keeps it low. Every call through the gateway is priced and recorded on a ledger, so an organization can see exactly what its automation does. Because the work runs as tools in a fixed order on local hardware, running it more often doesn't cost more. Refreshing hourly costs the same as refreshing monthly, so you choose the schedule for accuracy rather than budget.

Automation stops being an engineering ticket

Traditionally, recurring work belongs to whoever can write the cron job, so it gets set up once and then left alone. Whether it runs in Nucleus's cloud or on a Dispatch box, a workflow is configuration rather than code, and anyone with access can set one up.

from a laptop

Ask Claude

“Keep this site's translated cache fresh, check the sitemaps every six hours, and tell #web-ops when it fails.” Claude builds the workflow on the box over MCP and runs it once so you can check the output.

from a phone

The Dispatch app

Duplicate a workflow for a new customer, change its schedule, approve a large batch with one tap, and watch every run's steps, wherever you are.

from the channel

A Slack command

A slash command starts a workflow, and the result is posted back to the channel. The person typing never needs to know a workflow exists.

Governance doesn't loosen because more people can build automation. A workflow anyone can create still calls tools that only its owner is allowed to call, at that owner's rate limits, against that organization's budget, and every call carries that owner's name. Nucleus is why it's safe to let anyone build.

See it running

The same translation cache, 99.84% current instead of 23.8%.

The Flatten use case → What Dispatch is