How Proliferate works
How the pieces of Proliferate fit together.
Proliferate runs coding agents like Claude Code, Codex, OpenCode, and Cursor inside git-backed workspaces, then leaves a real diff to review before anything ships. This page is the map. Architecture goes deeper on each piece and the boundaries between them.
The short version
Every task happens in a workspace: a logical unit that pairs a repository and branch with one execution target, its chat history, sessions, and review state. Its files are materialized at that target — a checkout or worktree on your machine, or a worktree in a cloud sandbox — and that's what the agent reads, edits, and runs commands in. You open a workspace, pick an agent and a model, then send a prompt. The agent works, streams its progress back, and leaves a branch you review like any other diff.
Three layers make that work:
| Layer | What it is |
|---|---|
| Client (desktop or web) | The surface you look at: chats, terminals, files, git review, settings. |
| Control plane (server) | Owns product metadata and configuration: accounts, orgs, workspace records, authentication, credentials, billing, integrations, projections, and routing. |
| Runtime (AnyHarness) | Drives the agent on your machine or in a cloud sandbox, and is authoritative for runtime state: sessions, transcripts, terminals, and materialized workspaces. |
Where credentials live
The client never calls a model provider or a third-party tool itself. Integration credentials stay server-side; the agent reaches Linear, Notion, or Slack only through a Proliferate-owned gateway that proxies each call. Model auth depends on the route: the agent may sign in natively (your own Claude or Codex login), use a direct API key, or route through a managed gateway. See Architecture for the exact split.
Local or cloud, chosen up front
When you create a workspace you choose where it runs: a local checkout, a new worktree, or a cloud sandbox. All three use the same core chat, terminal, file, and review model, but their lifecycle, isolation, available integrations, and wake behavior differ. On your machine, one shared local runtime serves all your local workspaces; in the cloud, a single per-user sandbox backs every cloud workspace you own.
Everything is git-backed, so moving work between targets is a normal git operation, not a special feature: push the branch from one workspace, open or continue it somewhere else. See Local & cloud.
Agents run through their native harness
Proliferate doesn't reimplement Claude Code, Codex, or any other agent. It launches each through its own CLI or protocol and exposes the native behavior its adapter supports. Authentication, models, permissions, and advanced features still vary by agent and surface. See Agents for the supported list.
Where to go next
Architecture
Every component, the ownership boundaries, and where code executes per target.
Workspaces
Local checkouts, worktrees, and cloud sandboxes in day-to-day use.
Agents
Which coding agents Proliferate runs and how to configure them.
Authentication
How agents pick models and authenticate with providers.