Ask → browse → approve → save. A live computer view and a human review card appear right in chat, then the approved draft becomes an editable Space page. Enlarged for readability; idle time is trimmed and playback is accelerated.
Overview
OpenDots is a starting point for building your own agent workspace. Clone it, define your Dots, connect your services, and adapt the interface and tools to your needs.
A template, not a hosted product. You run the application and configure its infrastructure. The template is in early development; the Features section below describes what’s included and distinguishes local verification from connected-service testing.
Spaces
A Space is a home for working documents. Dots appear separately in navigation and can be granted access to multiple Spaces in their settings. Each Dot has a default destination for saved pages; existing installations retain their original Space access. Browse pages in a searchable library, switch between grid and list views, and organize documents as nested subpages. Open a page in a focused visual editor with formatting, slash commands, and undo/redo. Write directly, save a conversation as a page, or ask a specialist to create and revise content.
Pages stay in the local workspace database. Their conversations use CopilotKit Threads, with a separate conversation for each page and specialist. Page links connect the document workspace to Dot chat. Manual editing works before you configure conversation services. Autosave reports its progress, failed saves retain your draft, and revision checks prevent stale edits from overwriting newer content. Markdown source mode remains available.
Open a Space, navigate to its launch brief, ask Scout about the saved page, and continue in Dot chat. This recording uses live page chat and example launch content.
Specialist Dots
Give each Dot a name, role, instructions, and permitted tools. A researcher can investigate a topic; a writer can turn findings into a draft. Inspect their work and control what they can do.
Dot computers
Each Dot can have its own computer, using OpenBot’s container supervisor and computer service. Its browser profile and workspace files persist across stop/start. The Computer panel exposes browser control, human takeover, files, terminal output, and activity, with browser, file, and shell permissions set per Dot. The application keeps service credentials on the server and derives a different computer credential for each Dot.
See Computer setup to build the pinned services and connect your deployment. Computer tools require those services; an unconfigured template does not execute commands on your host.
Ask Scout to open a website, summarize it, save notes, and verify the file. Every computer action in this demo is requested through chat; CopilotKit tool renderers show the live browser, saved file, and terminal output inline.
Review before saving
Ask a Dot to show a draft before saving it. A CopilotKit human-in-the-loop card pauses the conversation for Approve & save or Decline. Approval creates the page in an authorized Space and returns a link; retries with the same draft recover that saved page. A changed draft needs a new review. The agent continues after your decision.
Connections
Give a Dot tools from any MCP server, such as email, calendar, GitHub, or your own services. Read-only tools run on their own. Any other tool pauses for an Approve & run card in chat, and only your approval runs it. You can turn each tool on or off per Dot. See Connections.
Connected apps
Dots can use Composio to work with the owner’s email, calendar, and other connected apps. Add COMPOSIO_API_KEY, restart OpenDots, then enable Connected apps in both the global settings and the Dot’s settings. Adding the API key alone does not grant access.
You can limit each Dot to toolkit slugs such as gmail or google_calendar. A blank toolkit list allows every toolkit in the Composio project. OpenDots identifies the owner with OWNER_ID, so keep that value stable after connecting accounts. When a tool needs authentication, the Dot returns a Connect link in chat for the owner to complete the provider’s sign-in flow.
These tools act on real accounts. Start with a read-only request and use a test account before trying write actions. See Composio setup for configuration, permissions, testing, and data-sharing details.
Text and calls
A continuous conversation keeps the Dot’s avatar and status above the messages, with text and call controls close at hand. Work updates, source links, and call receipts appear in the timeline; a side panel shows results or the agent’s computer.
Calls pair realtime speech with a separate compute agent, so the conversation can continue while longer work runs. Both use the same conversation context and tool permissions. The call screen includes a live timer, separate user and Dot captions, microphone mute, speaker mute, and a minimized view for continuing in chat. Voice needs separate provider configuration.
Connect, talk, mute, minimize, and return to chat. This is a silent screen capture of a real call, with waiting time trimmed and playback accelerated.
Slack
Mention a Dot through a managed Slack connection using Channels SDK, then continue in its thread. The integration follows OpenTag, with an explicit workspace/user allowlist and a selected specialist. See Slack setup to connect your deployment.
Bring your agents into Slack with Channels SDK. See the managed Channels documentation to connect them through CopilotKit Intelligence.
Architecture
AG-UI connects the agent to the interface
AG-UI carries streamed messages, tool calls, and agent state between the backend and CopilotKit components. Computer activity appears inline as the agent works; human-in-the-loop cards pause a tool call for your decision before it continues.
The template uses TanStack AI for model streaming and server-tool execution, CopilotKit’s React SDK and runtime, Intelligence for durable Threads, and Channels SDK for Slack. Pages, application metadata, and background-work state are stored separately from conversation history.
flowchart TB
Web["Web app: pages, Spaces, Dots, chat"] -->|AG-UI| Runtime[CopilotKit runtime]
Slack[Slack] <--> Managed[Managed channel connection]
Managed <--> Channels[Channels SDK]
Channels --> Agents[Specialist compute agents]
Runtime --> Agents
Agents --> AI[TanStack AI]
AI --> Provider[OpenAI-compatible model provider]
Runtime <--> Intelligence["Intelligence / Threads"]
Channels <--> Intelligence
Web <--> Speech[Realtime speech]
Speech --> Bridge[Compute bridge]
Bridge --> Agents
Agents --> Controls[Tool permissions]
Controls --> Computer[Isolated browser / workspace]
Agents --> Jobs[Background work]
Runtime --> Metadata[(Pages, Spaces, Dots, work metadata)]
You configure the Intelligence project, model provider, and channel connection for your deployment; calls also need a speech provider. Credentials stay on the server. Missing configuration should produce a clear setup state, and test fixtures should remain visibly separate from live integrations.
OpenMuse and OpenBot are code references for persistent work, agent computers, and execution controls. OpenDots can be adapted to your own workflows and deployment choices.
Get started
Use Node.js 24 and npm:
git clone https://github.com/CopilotKit/OpenDots.git
cd OpenDots
npm ci
cp .env.example .env
npm run dev
Open http://127.0.0.1:5173. You can create Spaces, write pages, and configure Dots before connecting services. To start chatting, connect an Intelligence project and a model.
Choose where to store conversations before connecting. For local Docker evaluation, follow Local Intelligence evaluation. For hosted Intelligence, sign in and select or create a project in the OpenDots folder:
npx copilotkit@latest login
npx copilotkit@latest project select
project select writes a project key to .env as CPK_INTELLIGENCE_API_KEY. Keep the generated CPK_TELEMETRY_ID with that key so SDK usage can resolve to your Intelligence account. Then add OPENAI_API_KEY and OPENAI_MODEL to .env and restart npm run dev.
Do not run copilotkit onboard in this folder. OpenDots already contains its CopilotKit integration, and onboarding adds a second, generic one.
See Setup for configuration, Slack, calls, the browser service, and Docker.
Data and privacy
Conversation messages, tool calls, and run events are persisted in the configured Intelligence deployment. The hosted setup sends them to CopilotKit’s cloud; the local evaluation setup uses Intelligence on your machine. OpenDots has no standalone SQLite conversation store. SQLite stores pages, workspace metadata, and thread bindings separately.
The configured model provider receives conversation context, including authorized page content and tool results. Local Intelligence does not make a remote model local: configure the app’s model and the Automatic Learning model separately. Public-web research sends queries and selected URLs to Parallel by default; set WEB_SEARCH_PROVIDER=disabled to disable those tools. When connected apps are enabled, tool arguments and results pass through Composio and the app provider. Speech and messaging integrations send data to their configured providers when used.
CopilotKit SDK telemetry collects usage metadata separately from conversation persistence. OpenDots also records bounded browser setup stages and first successful assistant activation with a random installation ID; the same opt-out flags disable this tracking and purge pending setup events. See browser setup telemetry. Runtime events are tagged OpenDots; see signup and usage tracking. Set COPILOTKIT_TELEMETRY_DISABLED=true or DO_NOT_TRACK=1 to opt out; see CopilotKit telemetry. Review CopilotKit’s privacy policy and the policies and retention settings of each service you configure. Installing or renewing local Intelligence still requires CopilotKit sign-in and internet access.
Features
| Area |
Included |
| Spaces and Specialist Dots |
Saved names, role instructions, and per-Dot research and memory permissions |
| Pages |
Searchable library, visual editor, slash commands, autosave, and revision checks |
| Conversations |
React SDK chat and Threads integration, page-specific conversations, and source links |
| Slack |
Managed Channels SDK declaration with workspace and user allowlists |
| Calls |
WebRTC speech, delegated compute, bounded sessions, hangup, and timeline receipts |
| Background work |
Scheduled server-side turns in their original conversation, with pause and retry controls |
| Browser |
Separate read-only public-page service with page capture and navigation limits |
| Dot computers |
Per-Dot browser profiles, files, shell, takeover, permissions, and action records through OpenBot |
| Connections |
Per-Dot MCP servers, Composio connected apps, per-tool access, and owner approval for MCP write actions |
| Appearance |
System, light, and dark themes; dark colors are derived at build time from the light styles |
| Memory |
User-managed preferences that permitted Dots can use |
| Automatic Learning |
Per-Dot Learning containers, conversation evidence routing, and published-skill delivery; see setup |
| Deployment |
Local Node setup and separate application/browser containers |
Scheduled tasks run in their original conversation. If a worker stops or its lease expires during a run, OpenDots marks that run Interrupted and waits for an explicit retry. Review its pages and computer actions, then use Retry after review when appropriate. Completed effects may already be present even when a run has no final result.
Local checks cover setup, persistence, permissions, SDK failure handling, and browser isolation. Automated tests use service fixtures. Live Intelligence, model responses, and page-context chat were verified on September 29, 2026. Live OpenBot computer browsing, file creation, shell verification, and file persistence across stop/start were also verified locally. Live Realtime speech, call controls, and receipt persistence were verified locally on September 30, 2026. Slack and spoken compute delegation still need connected-service verification. See recording notes for the demonstrated flows and limits.
Automatic Learning routing and skill delivery are configured locally. Cloud schedules, eligible-thread counts, and published-skill delivery still need connected-service verification. Skills require review and publication in Intelligence; existing conversations without a container are not enrolled retroactively.
This is a single-owner starting point. Shared editing, invitations, file uploads, and interactive page embeds are not included. Schedules are recurring instructions, not a complete goal or event-trigger system. Specialist Dots have separate roles and conversations; multi-Dot group conversations and automatic delegation are further work.
Extending the template
- Add identity, Space membership, and shared page editing for multi-user deployments.
- Add file attachments and richer page content.
- Add event triggers and a persistent responsibility lifecycle.
- Extend tools and approval flows for your own workflows.
- Add richer artifacts, connected-app context, and specialist coordination.
Contributing
See Contributing for development guidance and Security for reporting issues. Contributions should describe the workflow they enable, include verification evidence, and distinguish live integrations from fixtures.
References
License
MIT.
Public-web research
Parallel is selected by default in live research and Dot conversations. Ask a topic-only question to discover and read up to five sources, or supply URLs to extract them directly. Sources are saved with their links. A browser worker is not required for this research path; computer tools remain available for interactive work.
WEB_SEARCH_PROVIDER=browser preserves the existing URL-only browser reader, and WEB_SEARCH_PROVIDER=disabled disables these research tools. Workspace and Dot research permissions, pause and cancellation controls still apply. Sample research remains fictional and does not contact a provider.
Queries, requested URLs, a stable session identifier and the research objective are sent to https://search.parallel.ai/mcp. Memories and complete conversations are not automatically forwarded to Parallel. Model-selected objectives may still contain context from the conversation. The anonymous service is free for light use with provider-managed limits; set PARALLEL_API_KEY on the server for production or higher limits. Provider errors and empty results are reported rather than replaced with invented evidence. See Parallel Search MCP documentation.
![renovate[bot]](https://avatars.githubusercontent.com/in/2740?v=4)
评论