Field noteNov 17, 20266 min read1 source

What's new in OpenBot 1.0

In one minute

CopilotKit’s OpenBot 1.0 turns its self-hostable agent workspace into a broader desktop and team environment: users can bring compatible agents and models, schedule recurring work, and teach bots through reviewed task recordings. For builders, the release is notable because it combines more autonomous execution with explicit approval, policy, and audit controls.

What happened

CopilotKit announced OpenBot 1.0 on November 10, 2026. The company describes it as a modular, self-hostable agent workspace with built-in governance, available as a desktop app for Mac and Windows or for self-hosting. It says the desktop setup installs the backend on the user’s computer without requiring a terminal. OpenBot is open source under the MIT license and free for personal use; CopilotKit directs organizations to a separate OpenBot for Business offering.

The release broadens model and agent choice. During setup, users can sign in with Claude, ChatGPT, or xAI subscriptions, or choose Gemini, Unbiased AI, or an OpenAI-compatible endpoint, including Azure, Bedrock, Mistral, DeepSeek, Ollama, vLLM, or a custom endpoint. Users can switch the model and thinking effort for a conversation from the composer. Agents that speak AG-UI can be added as Bots by pasting an agent URL. CopilotKit names integrations built with Claude Managed Agents, Google ADK, LangChain, Mastra, Microsoft Agent Framework, Pydantic AI, Agno, LlamaIndex, and AWS Strands, among others. OpenDots, OpenFuse, and OpenDuet are now selectable interface skins.

OpenBot 1.0 adds several ways for bots to continue work beyond a single exchange. A bot can divide a task among up to five subagents working in parallel. Users can create routines in plain language, such as a weekly briefing, and then pause them, run them immediately, or inspect past runs. Bots can also monitor changing material, such as an inbox or shared document, and notify a user when attention is needed. To teach a task, a person records it while performing it in OpenBot’s browser and can narrate the steps. Automatic Learning turns the recording into a proposed skill; a person must review and approve it before any bot can use it.

The company also describes computer access and safeguards. A person’s bots share one computer by default; OpenBot for Business can give a bot an isolated computer. The bot’s screen is a browser the user can take over, and its shell can be watched live. The desktop app can grant read-only access to selected folders and asks first before editing. Bots ask before deleting, sending, sharing, or posting, including in routines; payment remains with the user. Settings allow custom approval rules. Business features include SSO and SCIM, named action boundaries checked before each action, a filterable and exportable audit trail, and organization-wide controls for models, MCP servers, network rules, and browser use. CopilotKit says the business edition can be deployed to Kubernetes using a Helm chart on EKS, AKS, GKE, or another cluster.

What the archive adds

New Runtime’s earlier coverage of OpenBot on October 27, 2026 described a source-available desktop app for running multiple persistent agents through one interface. The 1.0 announcement is a new release with additional capabilities and a wider product scope, rather than a restatement: it specifies model and agent choices, routines, reviewed task learning, and governance controls. The overlap makes this a new fact on a known story.

Trend signal

This release supports agent protocols become the interoperability layer. OpenBot says an agent that speaks AG-UI can be added by URL, while the workspace can also use a range of models and tools. That is a concrete example of a workspace positioning a protocol as the connection point between its interface and agents built elsewhere. The announcement describes compatibility, but does not establish how broadly or reliably those integrations work in production.

Why now

The timing is tied to the 1.0 release itself: CopilotKit says it is incorporating features requested by the community and making the workspace easier to start using and customize. The company also says it has passed 10,000 GitHub stars. That is a project milestone reported by the publisher, not evidence by itself of deployment scale or enterprise adoption.

Why it matters

For teams building agent workspaces, OpenBot 1.0 brings together choices that often sit in separate layers: the model, the agent implementation, persistent routines, browser and shell access, and controls over actions. That combination can make it easier to test an existing agent inside a shared interface rather than rebuild it for one vendor’s model or product. The capability to divide work among five subagents and run schedules also changes the operating question from “Can a bot answer?” to “How should a bot continue work, and who sees or approves the result?”

The announcement’s strongest practical idea is that autonomy and permissioning are designed together. A recorded task does not automatically become a reusable capability: someone reviews and approves the proposed skill. Sensitive actions trigger confirmation, while business administrators can set boundaries and inspect an audit trail. Builders can use this as a design checklist when adding recurring work or task learning to their own systems. The announcement does not show whether these controls are sufficient for a particular organization’s risk profile, but it makes their place in the workflow explicit.

Limits and uncertainty

The announcement is a product description from CopilotKit, not an independent evaluation. It does not provide comparative reliability, latency, or task-completion results for the new capabilities, nor does it quantify how well routines, monitoring, or recorded skills perform across different workflows. It names compatible agent frameworks and model endpoints but does not report integration-by-integration test results. The company describes approval rules, action boundaries, and audit features, but the announcement does not establish their effectiveness against misuse or failure in deployment. Business pricing is not stated in the announcement.

Skill to build now

If you are adding task recording to an internal agent, make approval part of the skill’s publication path rather than treating a successful demonstration as sufficient. OpenBot’s described flow offers a concrete sequence to try this week: record one repeatable browser task, turn it into a proposed skill, have an owner review it, and publish only after approval. For any recurring workflow that can send, delete, share, or post, define which actions require a person’s confirmation before enabling the schedule.

Reader takeaway

OpenBot 1.0 is not just a desktop wrapper for chat: CopilotKit is positioning it as a self-hostable workspace where agents can keep working, learn reviewed skills, and connect across protocols and models. The useful builder lesson is the pairing of those capabilities with explicit approval and governance paths. Whether the product delivers those controls reliably in real deployments remains to be tested.

Evidence / sources

  1. 1copilotkit.ai/blog/openbot-whats-new
All Field Notes

Next reads

Read next

Open the graph
  1. 01topicClaudeExplore the claude topic hub.
  2. 02topicGeminiExplore the gemini topic hub.
  3. 03archiveField NotesOpen the latest editorial analysis.
  4. 04sourcesSourcesSee the original sources this archive cites.
Who read this page?Machine requests, hidden until opened

Loading the privacy-safe route aggregate...

Open the JSON contract