Why Swarm

Efficiency is not only about making the interface lightweight. It is about reducing the amount of repeated inference, human supervision, state reconstruction, and coordination required to move from an idea to a trustworthy result.

Swarm approaches that problem with durable sessions, explicit plans, reusable context, provider choice, permissioned tools, and isolated worktrees for parallel changes. These are practical controls today, but they also create the foundation for deeper work on routing tasks to the right models, preserving useful state, and avoiding unnecessary computation.

This is a direction we are building toward, not a claim that every task already uses less model compute. The standard is measurable: more completed, reviewable work from the same hardware, energy, money, and human attention.

Swarm is not waiting for a distant future in which people disappear from the loop. People still set direction, supply judgment, understand consequences, and decide what deserves to ship. Agents can explore, implement, verify, and coordinate at a scale that would be difficult to manage one terminal at a time.

The platform should make that relationship better now: humans should be able to express intent clearly, inspect progress without reconstructing it, and intervene at meaningful decision points. AI should receive durable context, clear boundaries, and an environment designed for sustained work rather than isolated prompts.

More agent activity is not the goal. Better collaboration, better decisions, and more useful outcomes are.

Swarm's open-source runtime is local-first so your projects, durable sessions, tools, and credentials can remain in an environment you control. We do not want the work happening inside the app to become a behavioral data product, and we do not plan to collect project files or activity for Swarm-operated analytics.

The product signal we intend to keep is deliberately narrow: an anonymous registration event that tells us a Swarm was created, without recording an IP address or ingesting a stream of in-app behavioral analytics. That boundary is part of how we want the installed runtime to serve its users rather than observe them.

That statement is intentionally scoped. This website is hosted and uses third-party services, whose own data practices may apply. Configured AI providers and external tools receive the information needed to fulfill the requests you send them. Swarm Cloud is a managed service, so operating it safely will require infrastructure logs, health signals, security monitoring, and account and billing data. Those hosted responsibilities should be explained plainly, not blurred into claims about the local runtime. Current website practices are described in the Privacy Policy.

The immediate work is the Swarm open-source runtime: a dependable local platform for coordinating models, agents, tools, sessions, and repositories. It should remain useful on its own, understandable by the people running it, and open to the providers and workflows they choose.

The longer horizon is a cohesive AI-native computing experience, one in which the runtime, interface, system integration, and resource management are designed together. That future should not replace the open runtime or depend on removing people from the process. It should make the same principles easier to experience: efficient use of compute, durable collaboration, clear control, and technology that remains accountable to its users.

Swarm is being built for the work in front of us and for the scale of work coming next. The mission is to help people and AI do more with less waste, without giving up agency, privacy, or the ability to understand how the work gets done.