CopilotKit Open-Sources OpenDots: a Self-Hosted Answer to OpenAI's Dots That Runs Any Model and Owns Its Own Computer

When OpenAI turned Dots into the centerpiece of DevDay on September 29 — always-on agent personas that work in the background on 4,000 apps — the obvious question among developers was when someone would ship the open version. The answer came quickly: CopilotKit released OpenDots on October 1 under the MIT license, a self-hosted template for the same pattern, with source on GitHub at CopilotKit/OpenDots.

The pitch is structural. OpenAI's Dots are closed, tied to OpenAI's own models, and gated by plan and region. OpenDots runs on your own server, works with any OpenAI-compatible model, and brings any AG-UI agent to the party. Each Dot can be given its own computer — a browser, a file system and a terminal it operates independently — though computer access ships switched off by default.

The feature list tracks the closed original closely. Dots share Spaces and Pages, join live calls, plug into Slack through the Channels SDK, follow scheduled tasks, and accumulate skills through what CopilotKit calls Automatic Learning. A draft-approval flow stands out: by default a Dot prepares its work and waits for a human to approve before anything is saved or sent, a deny-by-default posture that most agent tools bolt on late rather than build in first.

Under the hood, OpenDots is assembled from CopilotKit's existing stack — the CopilotKit runtime, the AG-UI protocol and CopilotKit Intelligence — and runs on both web and mobile. The quickstart is the standard open-source routine: clone the repository, npm ci, fill in a .env file, npm run dev, then wire in AG-UI agents and enable the computer or Slack connectors only as needed. The team describes the current release as a single-owner starting point by design, not a multi-tenant product.

Traction was immediate. The repository was created on September 29, the day Dots launched, and had collected roughly 1,448 stars and 176 forks within the first days, trending on GitHub shortly after publication. The launch post is notably careful about claims: there are no SWE-bench or OSWorld numbers, and CopilotKit does not invent any, positioning OpenDots as a comparison on openness and portability rather than on closed-model leaderboards.

The strategic read is that the "persistent coworker" pattern is becoming infrastructure, and infrastructure tends toward open implementations. Dots, Meta's Muse and xAI's Grok Bot are all converging on the same shape: an agent with memory, a workspace, access to tools like Slack, and the initiative to keep working between conversations. Enterprises will adopt that shape only where they control the data and the model choice, which is exactly the gap OpenDots targets.

Whether it holds up is another matter. Self-hosting shifts the operational burden to the adopter, an always-on agent with its own terminal is a genuine security surface regardless of who ships it, and OpenAI can out-integrate a small team on app coverage indefinitely. But as a statement about where agent platforms are heading — draft approval as default, model-agnostic by construction, auditable because the code is public — OpenDots makes the open counter-argument at the moment the closed category is being defined.