Product history
Last updated 21 July 2026
A curated history of the meaningful changes to Harkster, grouped into product milestones and shown newest first.
20–21 July 2026
The public Harkster website was comprehensively rebuilt with new product and technical content. The MCP server then gained API-key authentication and tools for retrieving research-inbox tags and trade ideas, turning the earlier MCP prototype into an authenticated integration surface.
7–14 July 2026
Feed items and files were refactored to use the same workflow-and-step model, replacing separate processing pipelines with a more consistent architecture. The trade-idea consensus ordering was also adjusted to improve how combined ideas were presented.
2–25 June 2026
Harkster added trending themes built from research-inbox clustering workflows. New queues, workers, summaries, administration screens, and an API endpoint supported cluster creation and management, while post-release refinements improved trade ideas, theme presentation, and source handling.
19 May–1 June 2026
Trade ideas gained embeddings and a substantially redesigned extraction pipeline. The application introduced a consensus view and supporting modal so users could inspect how ideas from multiple research sources related to one another.
12–14 May 2026
Harkster began adding agent-based workflows and introduced its first Model Context Protocol server. This created a new path for exposing Harkster’s research data and capabilities to AI assistants and external agent workflows.
15–27 April 2026
“Attachments” were generalised into files and moved into a SQL-backed processing pipeline shared with feed items. Mailbox Q&A returned to SQL, watchlist publishing was restored, and migrations prepared existing conversations and questions for the newer research-inbox model.
3 March–14 April 2026
Attachment processing evolved into a job-based pipeline with grains, queues, dispatchers, and workers. Documents were converted into Markdown sections, chunked, processed in parallel, and linked through reusable metadata. Research inbox items and publishing were introduced, and SQL Server vector search was explored as part of the new retrieval architecture.
17–27 February 2026
A new Harkster API was created with API-key management and endpoints for trade ideas and source content. Attachment extraction also began using Azure AI Content Understanding, laying the groundwork for more structured analysis of documents beyond ordinary email text.
8 January–16 February 2026
Harkster added end-to-end feed management, feed items, attachment upload, publishing, and in-app feed views. The underlying grains were standardised around versioned SQL-backed stores, inbox rules were extended to feeds and providers, and saved and trash views joined the primary navigation.
3–23 December 2025
Mailbox tags, questions and answers, and cached data moved into grain-based or Cosmos-backed storage, with common queries centralised. Feed categories, providers, subscriptions, controllers, and early feed-management screens were rebuilt, bringing the original feed concept back into the modern product.
10–26 November 2025
The single-page application and its API were converted to Microsoft Entra identity and connected to the new Harkster kernel. Mailboxes and SignalR notifications were adapted to the new identity model, forwarding gained an address book, save and restore actions appeared in the mail header, and a refreshed website accompanied the release.
8 October–7 November 2025
Backend services were reshaped into daemon-style workers and the Microsoft Graph integration was rebuilt. A dedicated webhook API, dead-letter monitoring, worker health reporting, blacklists, and background maintenance improved resilience. Immutable message identifiers also enabled save, reply, and forward workflows.
23 July–17 September 2025
Maintenance controls were added for Graph notifications and mailbox messages before Harkster introduced its first watchlist experience. Subsequent work added unread-state handling, sentiment analysis and visualisation, and moved user and mailbox management into an administration application prepared for private-network deployment.
7 May–5 June 2025
Microsoft Graph notifications moved toward an Event Hubs-based architecture. Users gained in-product feedback and request controls, trade-idea pins, retroactive tag application, full-text trade-idea search, and currency-pair support, with delta queries improving incremental mailbox updates.
2–15 April 2025
A new question-and-answer assistant was added to search and refined through several rounds of testing. It retained recent mailbox searches, adopted the o3-mini model, and introduced a more compact interface for conversational research.
21–31 March 2025
Attachment storage and summarisation were redesigned so identical files and their summaries could be reused across multiple emails. Sequential processing and Graph notification deduplication reduced duplicate work, a CLI was added for migrations, and trade ideas became directly searchable.
26 February–19 March 2025
Harkster introduced its first Microsoft Graph mail integration, including subscription lifecycle monitoring, folder handling, logging, wildcard whitelist rules, and bulk whitelist upload. Authentication moved to Microsoft Entra, while Trading Economics data and search-ranking refinements strengthened the research experience.
28 November 2024–19 February 2025
Search filtering and layout were refined alongside repeated improvements to search-summary, catch-up, and event-risk prompts. The calendar’s quick summaries were reworked, application settings expanded, and the platform moved from .NET 8 to .NET 9.
31 October–26 November 2024
PDF summarisation was added and attachment content was incorporated into trade ideas, catch-up, email summaries, and search chunks. The event-risk calendar gained pre- and post-event summaries, while “co-pilot” was renamed “search summary” to better describe its role.
30 September–18 October 2024
The product expanded beyond the inbox with an event-risk calendar, tenant reporting and administration, mailbox catch-up summaries, and an initial chat interface backed by search. These additions began turning Harkster into a broader research workflow rather than a simple email reader.
2–26 September 2024
Harkster began extracting trade ideas from research and added asset-class metadata. Search prompts, inclusion filters, ranking, caching, chunk tagging, and spam controls improved discovery, while a health application and pipeline performance work made the service easier to operate.
20–31 August 2024
Cursor paging, timezone support, SignalR updates, mailbox-size tracking, mobile improvements, and an administration area made the inbox more operational. Search and the co-pilot were then added to the main application, completing the first integrated email-to-AI-search experience.
2–17 August 2024
The product shifted toward processing research delivered by email. Its data layer moved to tenant-oriented SQL storage, while new ingestion functions, a mail queue processor, mailbox services, expiry controls, and multi-recipient handling supported the first inbox rollout.
15–26 July 2024
Harkster gained a fuller ingestion pipeline with chunked feed items, vector embeddings, longer-context AI processing, and generated short and long summaries. Early search results and a “co-pilot” experience began turning ingested content into something users could query and interpret.
6 June–8 July 2024
The application was substantially reworked around a new Harkster AI experience. A dedicated inbox interface, marketing website, tagging system, and tag rules were introduced, followed by preparation for the first product release.
11 April–7 May 2024
The initial Harkster codebase established feed and feed-provider management, syndication parsing, text preprocessing and embeddings, Cosmos-backed storage, Azure search indexing, and an early processing pipeline. This was the first expression of the product as a system for ingesting, organising, and searching information.
