Replace OS vs Claude Desktop: a roster on the same local floor

Anthropic took local-first agents seriously as a shipping product, and the pattern they landed on was the right one: a sandboxed folder on your Mac, an explicit tool-use loop, careful gating around anything that could break the system, a small scheduler so the agent could keep working between sessions. None of it was glamorous, and that was the point — the chrome got out of the way of the work.

Replace OS is built on the same floor. Local-first storage, capability-scoped file system and shell per employee, inline approval for destructive actions, no remote sandbox VM in the loop. If you have already used Cowork the muscle memory carries — the same mental model of “agents work in a folder, you approve the risky bits” applies without modification.

What changes is everything above the floor.

Cowork is one Claude per sandbox. The sandbox can do a lot — long-context reasoning, multi-step tool use, scheduled returns to the same folder — but everything that happens inside it happens through one agent loop. If the task spans roles, the same agent puts on every hat.

Replace OS runs a hierarchy on the same floor. A CEO sits at the top and routes. Specialists report — an App Builder for code, a Marketing Strategist for launches, a QA Reviewer for checks. They run in parallel when the work allows, and the CEO converges the results before answering you. Peer review is a primitive: any employee can convene a moderated discussion with peers and have them critique each other's output before it ships.

Neither shape is wrong. One agent in a folder is the lightest possible step into local-first agents and the cleanest single-task experience. A roster in the same folder is what you reach for when the task has more than one role in it and you don't want to be the one threading the work between them.

Cowork's memory is scoped per project. Open the same folder twice and Claude picks up where it left off; open a different folder and it starts fresh. That maps cleanly to the sandbox mental model — the memory belongs to the place the work happened.

Replace OS scopes memory per employee. The App Builder remembers every repo it has scaffolded for you, across folders. The Marketing Strategist remembers every launch it has helped run. The memory follows the role, not the directory. This is the right shape when the same employee works across multiple projects — you want it to remember your writing style across every doc it writes, not relearn it every time you open a new folder.

The two models converge for the common case of one-role-one-project. They diverge when an employee is genuinely a generalist for a kind of task across many surfaces. A solo founder running both a SaaS repo and a writing folder will feel the difference within a week — the App Builder remembers the SaaS even when it is helping with a doc, and the Writer remembers the doc even when it is glancing at the repo.

Cowork runs on Claude. That is not a complaint — Claude is excellent at the long-context tool-use work the sandbox shape rewards. It is a constraint to know about: the agent's ceiling is the model's ceiling, and the model is fixed.

Replace OS treats the model as a per-employee choice. The runtime ships an LLMTransport layer that talks to OpenAI, Anthropic, local Ollama or vLLM, and any OpenAI-compatible endpoint. You can run the App Builder on a coding-focused model, the Marketing Strategist on a different one, and pin a local model for anything that should never leave the machine.

Procurement teams have asked for this — and so have people who just want to try a new model on one role without committing the whole stack. The migration story from Cowork is simple: leave Claude as the provider for any employee where the workflow already works, and only swap models on the roles that benefit from it.

Cowork's safety story is good. Tools are explicit, destructive actions prompt for confirmation, and the sandbox is real — Claude cannot reach outside the folder you gave it. Most of what we built into Replace OS's approval layer was lifted, in spirit, from the same patterns Anthropic was already shipping.

Where we extend it: approvals are per-employee, not per-app. The QA Reviewer can carry tighter approval rules than the App Builder; the Marketing Strategist might never need shell access at all. Tools are scoped at the catalog layer, so a non-engineer can hire an employee and trust the defaults without learning what the underlying primitives do.

The MCP story is similar. Both products talk to MCP servers, both surface the same approval contract for tool calls. The difference is mostly the surface area: Replace OS ships a curated default set of tools so a non-engineer doesn't have to wire any of it; Cowork leans more on the user to configure what they need.

Cowork is the cleanest fit when the work is one long task by one agent, in one folder, on one model. Long-context reasoning over a single document. A coding session focused on one repo. Research over a single corpus. The simpler the shape, the better the single-agent surface looks.

Replace OS is the cleanest fit when the work fans out across roles or surfaces. Operations work that touches code, copy, and customer ops in the same week. A founder running both a product and a marketing site. A team that wants per-role memory because the people doing the work specialize.

Both shapes are local-first, both gate the risky actions, both are calm desktop chrome. The choice is mostly between “one agent that goes deep” and “a team that distributes wide.”

Most of what carries over is the muscle memory. Tools are similar primitives; approvals work the same way; the file-system contract is the same; the model picker behaves like a power-user setting, not a daily thing to touch.

What is different at first: you don't pick which agent to ask. You DM the CEO, or the employee whose role obviously owns the task. The first day people often over-DM specialists and the second week they default to the CEO and let it route. Both work.

We don't have a one-click importer for Cowork projects — the data shapes are close but not identical, and the per-project memory in Cowork doesn't have a clean home in our per-employee model. The pragmatic path is to start fresh on Replace OS for the things that fit it best (multi-role work) and let Cowork keep the single-agent projects until you have a reason to move them.