Changelog

Smarter Threads, Smoother Flow

This release focused on the parts of Humm people touch all day: asking the next question, sharing work across a team, and moving through integrations without UI dead ends. You should see better follow-up prompts in threads, clearer ownership on shared work, and fewer places where the product gets in your way.

Suggested Next Questions

Humm now suggests follow-up questions directly under the latest AI response, so you can keep a thread moving without retyping the obvious next step. The suggestions combine fast defaults with model-generated ideas, which means they show up quickly and still feel specific to the analysis you just ran.

Shared Threads, Less Guesswork

Shared threads now show who started them in both the sidebar and the thread header, with clearer waiting states when Humm is expecting a specific teammate to respond. We also fixed a shared-thread follow-up issue that could leave a conversation in a bad state after the next question, which makes collaboration a lot less fragile.

Cleaner Connector Pages

Connector source pages now route to the right experience more consistently, so SaaS connectors show their Tools tab and sync connectors land in the sync-specific UI instead of the generic SaaS view. We also cleaned up the synced streams editor so it uses the screen better and feels less cramped when you're selecting a longer list of streams.

Bug Fixes

• Fixed a thread delete dialog that could freeze in the sidebar.

• Fixed shared-thread follow-ups that could corrupt thread state.

Data Sync Pipeline

SaaS data sync connectors are live. Connect sources like Gong, Stripe, HubSpot, and 15+ other SaaS tools, and Humm syncs their to a custom data warehouse automatically. This new style connector unlocks significant performance gains for the Humm agent and we expect it to become the default for most organizations.

Note: This feature is currently in beta. Please reach out if you want to test it: support@heyhumm.ai

Claude Desktop Native Connection

Humm now connects natively to Claude Desktop via MCP OAuth. Users add the URL, log in through Auth0, and they're connected.

@ Data Catalog References

Type @ in the chat input to search and reference tables, columns, and SaaS objects inline. Selected items render as typed blue badges with hierarchy breadcrumbs. The analyst agent automatically parses these references and injects focused schema context into its prompts, giving it precise knowledge of what you're asking about.

Additional Updates

• Chargebee connector fixes: Pagination, filters, response handling, and entity key unwrapping all fixed

• Citation rendering: Aligned citation footnotes with remark-gfm standard

• Agent tool loop fix: Stopped get_tool_details from entering a runaway loop

• Delete button styling: Delete buttons and menu items made legible

Humm MCP Server

Humm now speaks MCP. An MCP server embedded in the analyst service exposes Humm's data tools (`search_tools`, `get_tool_detail`, `execute_tool`, `search_context`) to Claude Cowork, Claude Code, Cursor, and any MCP-compatible agent. Authenticate with a Humm API key and start querying your data sources from your IDE.

File Uploads

Users can now drag-and-drop CSV and XLSX files directly into chat threads. Files are parsed server-side with inline previews showing column types and sample rows. The analyst agent can read the uploaded data via `read_cache`, making it easy to ask questions about your own spreadsheets.

Chart Optimizer V2

Charts got significantly smarter. Dual y-axis rendering automatically kicks in when two series differ by 10x+ in scale, 3+ mismatched series get split into separate sub-charts, and an LLM layer handles ambiguous cases. A quality scorer catches degenerate charts (single bar, flat scatter) before they reach the user. Validated across 80 test cases.

Clarification Cards

When the agent needs more information, it now presents structured clarification cards with single-select, multi-select, and free-text options in a wizard stepper UI. Cards persist in read-only mode after submission and survive page reloads.

Additional Updates

Agent Intelligence

• Dynamic summarization thresholds based on model context window (was hardcoded for 128K, now scales to 1M+ models)

• Meta tools for SaaS tool discovery let the agent find and use connector tools more effectively (currently in beta)

• Increased runaway detection threshold to 100 AI turns to support more in depth analysis

• JIRA_SEARCH_COMMENTS composite tool for searching across Jira issue comments

Frontend

• Intercom chat now available in the UI for direct support access

• Column filtering in InteractiveDataGrid

• Improved thread rename UX

• Horizontal scroll on code blocks in thread messages

• Fixed tool output field rendering and restyled DataAnalyzerOutput

More Ways to Connect Your Data

This week brought three new integrations and improvements to how you connect multiple accounts. We also fixed several bugs affecting authentication, emails, and scheduled reports.

Multi-Instance Connector Support

You can now connect multiple instances of the same integration. This is useful when you have separate accounts for different regions or business units.

When you add an instance label (like "US" or "EMEA"), Humm qualifies tool names automatically. Instead of generic CHARGEBEE_LIST_CUSTOMERS, you'll see CHARGEBEE_US_LIST_CUSTOMERS and CHARGEBEE_EMEA_LIST_CUSTOMERS in your agent's available actions.

To add an instance label: navigate to Data → Integrations → Sources, select your connector, and enter a label in the Instance Label field (20 characters max). If you only have one instance of a connector, no label is needed.

Three New Integrations

We added support for Zendesk, Amplitude, and Chargebee.

Zendesk brings in support tickets, users, organizations, groups, and satisfaction ratings. You can search tickets, retrieve comments, and analyze CSAT data.

Amplitude provides access to event tracking, cohorts, and user properties. Use this to join product analytics with your CRM or support data.

Chargebee exposes customers, subscriptions, invoices, transactions, plans, and items. Useful for analyzing billing patterns and subscription health.

All three are available in Data → Integrations → Catalog.

Improvements (3)

• Replaced browser confirmation dialogs with AlertDialog components across the app for cleaner design, consistent with the rest of the interface.

• Improved copy/paste behavior in the markdown input field.

• Strengthened English language enforcement across the agent system. All responses, reasoning text, and status updates now consistently appear in English, even when analyzing foreign-language data.

Bug Fixes (4)

• Fixed broken login button on unsubscribe page — it was pointing to /login instead of /auth/login, causing 404 errors.

• Fixed deep link preservation during authentication. When you navigate to a specific page while logged out, you now land on that page after signing in — not the homepage.

• Fixed scheduled report email links, timestamps, and display names. Email links now point to the correct URL, timestamps no longer show future times for past reports, and report detail pages show the schedule name instead of internal command names.

Image Input, Command Creation, and 100x Faster Thread Loading

This release introduces multimodal chat with image support, the ability to turn conversations into reusable commands, and major performance improvements across the platform. Users can now paste screenshots directly into queries, create custom slash commands from any thread, and experience dramatically faster load times.

Image and Screenshot Support

You can now attach images and screenshots to your analyst queries. Paste from clipboard, drag-and-drop, or click the attach button to send up to 5 images alongside text. Images are compressed client-side (max 1568px, JPEG 0.92 quality, 750KB) and appear as thumbnails in chat history.

To use: Click the attach button in the chat input, paste images directly (Cmd/Ctrl+V), or drag image files into the input area. Images work with all multimodal-capable models.

Create Reusable Commands from Threads

Turn any completed thread conversation into a reusable slash command. Click "Create Command" in the thread toolbar, name your command, and the system extracts the prompt pattern using an LLM. Your new command appears in the slash menu for future use.

To use: Open a thread with a conversation pattern you want to reuse, click the "Create Command" button in the toolbar, enter a unique name, and wait for generation to complete (typically 5-10 seconds). The new command is immediately available in your organization's slash menu.

Commands inherit the structure and intent of the original conversation, making it easy to codify repeatable workflows.

Manual Report Triggering

Scheduled reports can now be triggered manually via the "Run Now" button. Triggering executes in your permission context. The button disables while a run is in progress.

To use: Open the Reports inbox, locate the schedule you want to run, and click the "Run Now" button. The interface refreshes automatically when the run completes.

Thread List Performance

Thread list loading is now ~100x faster. Warm response time dropped from ~1,640ms to ~15ms.

Chart Datapoint Limits and Axis Fixes

Charts now enforce per-type datapoint limits to prevent unusable visualizations: 30 bars, 10 pie slices, 500 scatter points, 200 line/area points. When exceeded, the agent receives actionable feedback (aggregation suggestions, alternative chart types).

Scatter chart axis domains also fixed—no more wasted space with unnecessary negative ranges or buffer beyond the last tick. Y-axes crossing zero always include 0 with even spacing.

Improvements (4)

• Copy button now appears on all AI messages, not just the last one

• Share button replaced with dropdown: Copy URL, Share with org, Make Private

• Failed report runs show friendly error message with better contrast

• Removed summary truncation on scheduled reports

• Added Drive read scope to BigQuery connector for Google Sheets-backed tables

Bug Fixes (3)

• Fixed 405 error on report run deletion

• Prevented false quality warnings when final question passes validation

• Fixed table citations rendering with superscripts and links

• Fixed system threads getting stuck on ask_human prompts

• Fixed Chargebee tool parameter mapping to API format

• Empty SQL query results now return success instead of error

Authentication, Performance, and Platform Reliability

This release focuses on session stability, platform performance, and infrastructure resilience. Users will notice smoother authentication experiences, faster API responses, and more reliable data operations across PostgreSQL and SaaS integrations.

Authentication and Session Management

Users now stay logged in for 24 hours from their last activity, instead of being forced to re-authenticate after 30 minutes.

Improvements (8)

• PostgreSQL data sources can now execute SQL queries via the analyst agent (previously failed with "Unsupported data source type" despite a fully functional connector)

• Pendo connector auto-corrects common agent mistakes (merged stage splitting, source reordering, sort normalization, aggregator fixes, timeSeries nesting, unicode escape decoding) instead of failing with QueryError

• HubSpot tools consolidated into unified HUBSPOT_READ_RECORDS with smart routing (GET for single IDs preserving native associations, batch POST + associations for multiple IDs, auto-chunking >100 IDs into parallel calls)

Bug Fixes (2)

• Fixed citation anchor links opening new tabs instead of scrolling within reports — internal #citation-N links now preserve scroll-to behavior while external links still open in new tabs

• Clarified Jira JIRA_GET_ISSUE_DESC tool to explicitly warn that parameter is issue_key, not issueIdOrKey (raw API name), reducing validation errors from agent using API documentation naming