Local
Suitable workloads can remain on user-controlled hardware, using local models, deterministic tools, and LF's own Project/Memory/workflow state.
Red Rocket's local-first approach treats privacy, control, recurring cost, latency, resilience, provider portability, and user-owned infrastructure as architecture decisions rather than marketing footnotes.
Some workloads benefit from staying on infrastructure controlled by the user or organization. Others genuinely benefit from remote reasoning, specialized services, or larger compute. The useful question is not “local or cloud?” in the abstract. It is “where should this particular work run, under whose account, and why?”
LF's current architecture is designed so that local execution and supported provider APIs can participate in the same environment without moving the durable Project, Memory, Worker, permissions, or authority model into the provider.
The preferred posture can vary by customer, workload, hardware, privacy requirement, and desired capability.
Suitable workloads can remain on user-controlled hardware, using local models, deterministic tools, and LF's own Project/Memory/workflow state.
Keep routine or private work local while authorized remote reasoning or connected services handle the steps that genuinely benefit from them.
Supported provider reasoning can use user/customer-controlled credentials while LF remains the owner of the Project, Worker, permissions, tools, and history.
Local execution is not automatically better. It becomes valuable when it solves a real ownership, privacy, availability, cost, latency, or resilience problem.
LF separates inference identity from operational identity. A persistent Worker can keep its role, permissions, history, Mission continuity, and Project context while an authorized task uses a different model, provider, or reasoning effort.
That architecture makes model competition useful to the customer. A stronger, cheaper, more private, or better-specialized resource can be adopted without rebuilding the surrounding workflow.
Provider/model/effort and usage receipts remain inspectable so modularity does not become an invisible routing black box.
A suitable routing decision can consider whether the work should stay local, whether a specialist tool is better than a language model, whether a provider call is authorized, whether the local machine has capacity, and whether the task genuinely merits stronger remote reasoning.
LF remains the control plane around that choice: durable Project state, Memory, permissions, authority, tools, and receipts do not have to migrate simply because inference is remote.
GPU memory, CPU time, system RAM, disk I/O, startup cost, and network capacity are finite.
That is why Red Rocket treats resource awareness as part of orchestration. A model should not load simply because a task exists. The system needs to understand what the job requires, what is already reserved, what can coexist, and what should wait.
The same principle applies to provider usage: a remote call should not be the automatic answer to every problem merely because an API is available.
Local processing does not make a system automatically secure. Security still depends on device configuration, access controls, software, storage, networks, connected services, credentials, and user behavior.
Likewise, some capabilities inherently require internet-connected services. Local-first means unnecessary external dependency is not the default architectural assumption.