OpenDots builds persistent AI workspaces with approvals

What this is about

OpenDots is an open-source template for persistent AI workspaces. Teams can configure specialized agents called Dots with their own roles, tools, document spaces, and optional computers. The CopilotKit repository appeared on September 29, 2026 and uses the MIT license. It is not aimed at people who merely want to open a chatbot. Its audience is developers and internal platform teams that want to operate and customize their own agent workspace.

The approach brings together interfaces that are often separate: chat, a page editor, background work, browser control, voice, Slack, and MCP connections. The built-in approval steps are particularly relevant. Tools that write or otherwise cause effects are intended to pause until a human approves the specific call.

What OpenDots actually does

Users create Spaces for documents and configure Dots with a name, role, instructions, and permitted tools. A research Dot can collect sources; a writing Dot can turn them into a draft. Pages are stored in a local workspace database, and the template includes subpages, search, a visual editor, and a Markdown view.

For more demanding tasks, each Dot can receive its own isolated computer based on OpenBot. Browser profiles and workspace files persist across restarts. The interface exposes the browser, files, terminal output, and activity logs. Permissions can be set per Dot for MCP tools: read-only calls may run directly, while other calls require an approval card. Connections through Composio, voice conversations, and a Slack channel are also supported.

OpenDots is explicitly a template rather than a fully managed service. The web application runs locally with Node.js 24, but real conversations require a CopilotKit Intelligence deployment and a model provider. Computers, voice, Slack, and connected accounts require additional services and credentials.

Why it matters

A capable language model is not enough for persistent agents. Teams need traceable state, separated permissions, resumable work, and a place for results. OpenDots exposes this layer as inspectable source code rather than only as a product demonstration. A team can therefore tailor the interface, roles, and tools to its own workflow.

At the same time, the template reveals how many systems participate in such a workspace. Depending on configuration, conversation histories reside in CopilotKit Intelligence, model context goes to the selected model provider, connected apps may route data through Composio, and voice requires another provider. Self-hosting the web interface therefore does not automatically mean that every piece of content remains exclusively on the user's server.

In plain language

OpenDots resembles a small office building. Every worker has a job description, selected keys, and a separate desk. The worker may enter the archive alone to read a public document, but must submit a form for signature before sending an email or changing a file. The building helps with order and control, but it does not replace facility management or clear security rules.

A practical example

A five-person communications team wants to prepare a weekly market brief. It creates a research Dot with read-only access to public websites and a writing Dot with access to one Space. Every Monday, the first Dot gathers ten sources and saves notes. The second produces an 800-word draft, displays it in an approval card, and saves it only after confirmation.

The team initially tests with a separate account and fictional project data. External write tools remain disabled. After four weeks, it reviews 40 citations, failed actions, runtime, and model cost. Only when logs are understandable and rejection works reliably does it connect one real data source with minimal permissions.

Scope and limits

• OpenDots is in alpha. Setup, interfaces, and features can change quickly, and critical workflows have no long-term stability guarantee.
• The template consists of several services. Intelligence, the model provider, computer runtime, Composio, voice, and Slack have separate costs, outages, and privacy terms.
• Human approvals reduce risk but do not prevent misleading drafts, manipulated websites, or overly broad permissions. Results and tool calls still need review.
• Self-hosting is layered: pages and metadata can remain local, while conversations and tool data may pass through other systems depending on configuration.

A sensible first test stays small: one Dot, one Space, read-only tools, and synthetic data. Teams should enable computer access or real business accounts only after that.

SEO & GEO keywords

OpenDots, CopilotKit, persistent AI agents, agent workspace, human in the loop, MCP, OpenBot, AG-UI, self-hosted AI, Slack agents, browser automation, MIT license