UFO launches an open-source agent platform for whole teams

Why it matters

UFO's bet is that teams will adopt agents through a shared runtime that can execute work across company tools, not just through separate chatbots. Its self-hosted option and documented execution boundaries make deployment and data control central to that bet.

UFO, the San Francisco startup founded by Marshall Kiely, announced its hosted and open-source agent platform in a September 27th thread on X. The product gives teams a shared workspace for assigning work to agents through chat, the web, and a terminal. The company pitches it as infrastructure for coding and broader office tasks.

Kiely's background spans finance and machine learning: Y Combinator's company profile lists him as UFO's founder and says he previously led finance at Perplexity, worked in quant finance, helped found the AI engineering team at Notable Health, and worked on machine learning at Mixpanel. YC also lists a Princeton PhD in computational biology and first-author papers in Nature and Science. The profile identifies UFO as a Fall 2026 company with a two-person team. That background spans technical research and a product company's finance operation, while UFO's pitch covers software work and internal company workflows through one shared system.

UFO's launch thread says teams can use its agents through Slack, iMessage, the web, and a terminal. The product's documentation describes workflows that include making a code change through to a pull request, investigating an incident, analyzing data, and building an internal application. The launch post also lists website and custom-app creation, automations, office documents, pull-request review, and agent swarms. Those are company-described capabilities, not evidence that the system has demonstrated reliable results across each task.

The technical distinction is the runtime underneath those interfaces. UFO's public repository says the same code powers its hosted service and can be run on a customer's own infrastructure with customer-supplied model keys. In self-hosted mode, a basic installation can use SQLite and local files; the project also documents deployments with Postgres, S3, and Redis. Its Apache 2.0 license permits commercial and private use, while an extension system lets developers add tools, connectors, model providers, and other components.

The repository describes durable agent turns, workspace-scoped data, and account access that members grant to agents through chat. It also distinguishes between two execution modes: agents can operate on the connected user's machine or run in a configured sandbox. In the default local sandbox, the code says writes are confined but reads can reach the host machine and network egress is not enforced by the kernel; it recommends Docker or E2B for untrusted input or multi-workspace deployments. That detail is more consequential for teams connecting agents to company files and tools than the broad promise of an agent operating system.

Hosted convenience carries a different data boundary. UFO's privacy policy says that when customers connect services such as Google Workspace or Slack, UFO can receive the data permitted by those grants and send request-related content to model providers. The company's subprocessor list names infrastructure, model, search, and tool vendors used to operate the service, including AWS, Anthropic, OpenAI, Perplexity, E2B, and Slack. The policy says UFO does not use Slack or Google user data to train general-purpose language models. Teams choosing the hosted option therefore rely on UFO's access controls and vendor arrangements. Self-hosting changes where the runtime operates but still requires model credentials and an operator to manage the deployment.

UFO is entering a market where messaging apps are becoming a distribution channel for third-party agents. Slack's August 20th announcement is one example.