Open Assistant shipped eight releases in July, advancing from v1.2.9 to v1.3.6. The highlight of the month is the introduction of Model Context Protocol (MCP) server integration and a Plugin Creator agent, alongside significant improvements to Google Workspace, Outlook, sub-task execution, and the web interface.
MCP Server Integration (v1.3.6)
Open Assistant now supports integrating remote MCP servers as agents. This enables extending the assistant's functionality with external tools and services through a standardized protocol.
Key capabilities:
- Register MCP servers from the Settings interface and trigger their tools using custom intent keywords
- Multiple authentication methods: no auth, static HTTP headers, or OAuth 2.1 with PKCE flow
- OAuth integration with automatic token refresh and dynamic client registration
- MCP servers appear as distinct agent cards with transport badges, tool listings, and authorization status
- Edit server names and keywords, refresh tools, update credentials, and manage connections from the UI
- Persistent skill selection: once triggered by a keyword, MCP tools remain available throughout the conversation
On the technical side, the integration supports HTTP transport with flexible authentication, encrypted storage for sensitive credentials, automatic OAuth token refresh, and lazy loading to keep the application stable even when MCP packages or servers are unavailable.
Plugin Creator Agent (v1.3.0)
A new agent that installs and configures plugins directly from API specifications. Feed it an OpenAPI/Swagger URL or paste a plugin JSON definition, and it handles installation and connectivity verification automatically. It supports both OpenAPI 3.x and Swagger 2.0 formats.
The Plugin Creator is disabled by default and can be enabled in Settings > Agents. Combined with the Browser and Web Search agents, it enables a complete search, inspect, install, and verify workflow.
Recipe Visualization (v1.3.0)
Scheduled jobs now include a "Graph" button that displays a visual flow diagram of recipe execution. The diagram shows the complete data flow from trigger through steps to delivery, with color-coded variable tracking. It works with both multi-step recipes and legacy single-step jobs, and includes collapsible views for long prompt templates.
Rich Markdown Rendering (v1.3.5)
Assistant responses in the web interface now display formatted markdown instead of plain text. This includes full GitHub Flavored Markdown support: bold, italics, headings, code blocks, tables, lists, blockquotes, links, and strikethrough. All rendered HTML is sanitized to prevent XSS vulnerabilities. User messages and error messages continue to display as plain text for clarity.
Sub-Task Execution Improvements
Two releases addressed sub-task reliability and capability:
v1.3.5 fixed critical issues where asynchronous sub-tasks would falsely report as timed out while still running in the background. Results from completed sub-tasks are now properly collected and surfaced even if they finish after the coordinator moves on. The UI no longer blocks when background sub-tasks are running. Technical changes include a dedicated thread pool for LLM operations, concurrency limits for sub-task fan-out, and an increased default wait timeout from 300s to 900s.
v1.3.4 introduced skill pinning for sub-tasks, allowing them to be assigned to specific specialist agents by name. It also added task persistence: async task execution is now tracked and stored in the database throughout the task lifecycle, providing better visibility into background operations.
Google Workspace Editing (v1.3.2)
Non-destructive editing tools for Google Docs, Sheets, and Slides:
- Google Docs: find-and-replace functionality for targeted text updates without replacing entire document content
- Google Sheets: ability to clear cell ranges while preserving formatting
- Google Slides: create new slides with optional title and body text, find-and-replace across presentations, and insert text boxes onto specific slides
Bug fixes include correct retrieval of speaker notes and proper return of slide IDs in the Google Slides reading tool.
Outlook Email Optimization (v1.3.3)
A new summary_mode parameter for outlook_read_emails and outlook_search_emails returns emails in a compact, LLM-friendly format. Email bodies are automatically converted from HTML to plain text, reducing payload size by approximately 90-95% (from ~22 KB to ~1-2 KB per email). This significantly improves processing speed and reduces token usage when working with email data.
Array Parameters for Plugins (v1.2.9)
Plugin endpoints now support native array parameters, enabling seamless integration with APIs that expect JSON arrays in request bodies. Array parameters can be defined with a specified element type (string, integer, number, or boolean), with automatic JSON Schema generation and native serialization. This enhancement is fully backward compatible.
Mobile Responsiveness
The web interface received continued mobile improvements across several releases:
- v1.3.0: Monitoring > Jobs tab adapts to narrow screens with stacked card layouts, and graph visualizations scale appropriately
- v1.3.1: Artifacts page tables display as stacked cards on smaller screens; Settings > Agents tab adapts with vertically stacked headers and full-width buttons
- v1.2.9: Navigation bar displays an abbreviated "OA" brand text on mobile devices across all pages
Messaging Notifications Fix (v1.3.1)
Fixed an issue where the notify_owner coordinator tool would not appear when only Slack or only WhatsApp was configured. Automatic channel detection now works properly, preferring WhatsApp when enabled and falling back to Slack.
Token Consumption Tracking (v1.3.6)
Enhanced token usage calculation by capturing consumption data directly from the LLM client, providing more accurate usage metrics.
Background Task Nudge Limit (v1.3.6)
Fixed an issue where the system would repeatedly nudge the LLM about running background tasks indefinitely. The system now limits these nudges to 2 attempts, allowing the LLM to provide a final response while tasks continue in the background.
Get Started
Ready to try these features? Create an Open Assistant instance or learn more at open-assistant.org.