Honcho

Most organizations sit on a fragmented content estate: a CMS for published material, a wiki for internal notes, feeds and crawls for outside coverage, archives in cloud storage, and structured records behind internal APIs. There’s no unified way to query across all of it, and no clean way to push the consolidated view back out to the tools and systems that need it.

Honcho closes that gap. Ingest documents, feeds, API responses, and bulk archives into a single full-text index, then expose the consolidated content back out through GraphQL, REST, MCP, and replication endpoints. AI assistants are one of many consumers of the same data.

• Ingest from anywhere WordPress and other CMSes, feeds, internal APIs, raw HTML/JSON/XML, bulk archives, and pluggable connectors, all into the same index.
• Query across everything Full-text Lucene queries, faceted filters, structured fragment search, and configurable relevance over the consolidated corpus.
• Emit anywhere GraphQL, REST, MCP, feeds, sitemaps, and replication. Bring your own frontend, mobile app, AI assistant, or downstream system.

Content flows in from many directions: bulk importers for CMS migrations and document archives, pull replication from WordPress and other CMS platforms, feed crawling for ongoing sources, an extract API with declarative rules for structured data, pluggable Java connectors for custom integrations, and AI assistants that save articles and notes on your behalf. Tens of thousands of articles, hundreds of feeds, decades of archive: everything lands in the same full-text index.

Once indexed, the consolidated corpus is queryable through whatever protocol fits your stack. GraphQL for flexible structured queries, REST for JSON, feed, and sitemap output, and MCP for AI clients. Boolean search, faceted filters, time-range constraints, structured fragment lookups, and configurable relevance. The full power of Lucene is exposed through clean APIs. Push the same content back out through replication to other Honcho instances or downstream systems.

The same data is also reachable through human and AI surfaces: a web Console for operations (sources, users, tokens, metrics, Editions config), a built-in Copilot chat interface in the browser, and an MCP server that lets Claude, Claude Code, and other AI clients query the index directly. Use whichever fits the task.

On top of that foundation, you can ask “how have expert views on China’s economy evolved over the past decade?” and get an answer synthesized across thousands of indexed documents, threading together analysis, tracking how positions shifted, surfacing connections that no keyword search would find. Combine results from your library with live web content, search connectors, and internal APIs to compare, verify, and fill gaps.

It supports multiple users and groups, so a team can share a common pool of content and collaborate through their AI assistants. An analyst can summarize a report and share it to the team’s feed. A writer can research across the entire corpus and send findings to colleagues. A shared knowledge base builds organically without anyone copying and pasting links around.

Honcho is also a persistence layer for AI workflows. Agent apps that summarize articles, monitor topics, or produce analysis can write their output back to Honcho, where it becomes searchable alongside everything else. The content you curate and the content your tools produce all live in one place.

Most importantly, Honcho gets smarter as you use it. Every entry you save, every note you add, every article you mark as important becomes a curated relevance signal that shapes future answers. Rather than stuffing the AI’s context window with whatever matches the keywords, Honcho prioritizes your own research and annotations, so the AI answers in the context of what you already know and think, not just what’s in the library.

• Aggregate Crawl feeds and HTML sources on configurable schedules. A declarative rules engine lets you define per-site extraction logic (CSS selectors, field mappings, timestamp parsing, fallback chains) in a config file instead of code. Built-in deduplication and change detection keep the index clean.
• Ingest Content doesn’t just come from feeds. Pull replication syncs from WordPress and other CMS platforms with incremental updates. A pluggable connector API supports custom Java connectors dropped in as JARs. AI assistants save articles and notes via MCP. A save-URL endpoint supports bookmarklets and mobile shortcuts. An extract API accepts raw JSON, HTML, or XML and runs it through extraction rules. Bulk importers handle CMS migrations. Everything lands in the same index.
• Index Full-text search powered by Lucene with a configurable text analysis pipeline. Boolean queries, time-range filters, tag and topic facets, custom numeric fields, and configurable relevance scoring.
• Serve GraphQL, REST, and MCP (Model Context Protocol) endpoints for search, retrieval, and content distribution. No rendering opinions. Bring your own frontend, feed reader, mobile app, or AI assistant.
• Replicate Push and pull replication between instances and from external CMS platforms. WordPress connector with incremental sync, pluggable connector API for custom sources. Distribute content across organizational boundaries with topic-based routing.

• Search is built in, not bolted on The Lucene index is the primary read path. Content is indexed at write time with no sync lag. The query API exposes the full power of Lucene: phrase queries, field-scoped search, boolean composition, range-boosted relevance, and custom numeric fields for domain-specific ranking.
• Structured content fragments Content is modeled as typed fragments: paragraphs, headings, pull quotes, captions, recipe steps. Each fragment type is indexed separately, so you can search within specific block types: find all entries where a heading contains earnings guidance, or where a pull quote mentions supply chain.
• Declarative content extraction A rules DSL lets you define how to extract entries from any JSON, HTML, or XML source (selectors, extractors, transforms, timestamp parsers) without writing code. Rules compose via config layering: define a base for a feed or a WordPress connector, then override only the fields that differ per source. You can develop rules conversationally through an AI assistant: paste in your raw content, iterate until the extraction is right, then save them.
• One API surface, three protocols The same queries, the same writes, and the same content are exposed through GraphQL, REST, and MCP. Build a frontend in GraphQL, integrate a downstream pipeline via REST, and let an AI client drive the same operations through MCP, without maintaining three different integration layers.
• One system, not six Each service in the typical stack is another deployment, another set of credentials, another failure mode, another thing to keep in sync. Honcho’s tight integration eliminates the boundaries where things break.
• AI surfaces share the same foundation When you do want AI on top, natural-language queries translate to structured Lucene, search connectors reach live external data (SEC filings, court opinions, academic papers, economic indicators), the built-in Copilot adds a dynamic planner for multi-step analysis, and Editions turn the pipeline into an automated publication. All of it reads and writes through the same index your GraphQL and REST clients use.
• Curated context, not brute-force retrieval Standard RAG pipelines start from zero every query: keyword match, top-K results, hope for the best. Honcho builds a curated relevance layer from user engagement: every save, annotation, and boost is a pre-computed signal that no retrieval-time ranking can replicate. The AI answers in the context of what the user already knows and thinks (their notes, their perspective, their research trail), not just whatever matched the search terms.

• Runtime Java on Jetty with Jakarta Servlet.
• Search engine Lucene with unified numeric fields, near-real-time search with searcher warm-up, and proper FILTER vs MUST clause handling.
• API layer GraphQL for flexible queries, REST for JSON, Feed, and Sitemap output, and MCP (Model Context Protocol) for AI assistant integration.
• Data model Protocol Buffers for internal data model and wire format, with a purpose-built JSON encoder for browser and API clients.
• Instrumentation Dropwizard Metrics on every significant operation: search latency, indexing throughput, crawl rates, storage I/O.

• GraphQL & REST APIs The programmatic foundation. GraphQL for flexible structured queries with field selection; REST for JSON, feed, and sitemap output. Every read path the Console, Copilot, and MCP surfaces use is available to your own code on the same terms. Bring your own frontend, mobile app, newsletter generator, or downstream pipeline.
• Console The operational backbone. Source management and crawl health, full-text entry browser and editor, users and groups, collections, digests and Editions configuration, API tokens, LLM usage, metrics, account and preferences. The place you go to set things up, audit what the system is doing, and fix sources that aren’t crawling cleanly.
• Copilot Turnkey conversational surface in the browser. Zero setup. Just log in. Supports any LLM provider behind an OpenAI-compatible API in addition to Claude, including local models. Server-side orchestration is more token-efficient than MCP for heavy retrieval, because results are filtered before they reach the model instead of round-tripping tool calls. Adds chat history with forking, document upload, group collaboration, the dynamic planner, and Editions.
• MCP server Connect Claude, Claude Code, or any MCP-compatible client and let your client’s LLM do the work. Search the library, query external connectors, fetch URLs, search the web, create and update entries, save and recall memories, all as native tool calls. Works with claude.ai connectors (Pro/Max/Team/Enterprise), Claude Code, Claude Desktop, Cursor, VS Code, JetBrains, the Claude API, and anything else that speaks MCP.

Source management, favorites, group messaging, and other operational features stay in the admin UI and Copilot. MCP is intentionally focused on querying and writing content, not managing the system.

• Your client’s LLM does the work Claude (or whichever model your MCP client uses) does the synthesis, summarization, comparison, and planning natively, the same way it handles any other tool surface. Connectors, web search, URL fetch, and cross-session memory are all available as tool calls, so the model can plan multi-step analysis without server-side orchestration.
• Read and write AI assistants can search and retrieve content, but also create entries, update metadata, manage tags, and curate the index, all through the same MCP interface.
• External data connectors Query SEC EDGAR, FRED, CourtListener, OpenAlex, arXiv, GDELT, and other specialized sources directly from your AI assistant. Results are returned alongside your library content.
• Web + URL fetch Search the web and fetch URLs to compare external content with your library, fact-check claims, or fill coverage gaps, all within the same conversation.
• Cross-session memory Assistants can save and recall notes across conversations. Say “remember that the client prefers weekly reports on Mondays,” and it’s there next time you ask.
• Enterprise-ready auth OAuth 2.1 with PKCE, RFC 8707 resource indicators, RFC 9728 protected resource metadata, and dynamic client registration. Each user authenticates individually, so Claude only sees the sources that user is permitted to see. Team/Enterprise org owners install Honcho once as a custom connector; members connect with their own credentials.

• Zero setup No MCP client to install, no connector to register, no tokens to manage. Users log into Honcho and start chatting. For non-technical teammates, this is a meaningful difference.
• Bring any LLM provider Copilot works with Claude, and with any model behind an OpenAI-compatible API: OpenAI, Azure, Google, and locally hosted models. Users configure their own API key and model preference; usage is tracked per-user with full transparency.
• Token-efficient orchestration Server-side retrieval and ranking happen before results reach the model. For heavy library searches, that’s significantly fewer tokens than an MCP round-trip that has to ship every tool definition and intermediate result through the model. The dynamic planner can chain dozens of searches against connectors, web, and the library without burning context on tool plumbing.
• Dynamic planner Auto-executes multi-step analysis across sources: dossiers, source comparisons, gap analysis, trend tracking, briefings, fact triangulation. Source attribution on every claim.
• Document upload Upload PDFs and other documents directly into the chat for analysis, summarization, and indexing.
• Chat history & forks Conversations are saved and resumable. Fork a conversation to explore a tangent without losing your place.
• Group collaboration Share entries to a group’s feed, write summaries to the team’s research workspace, follow what colleagues are saving.
• Style guide Administrators set a system-wide editorial style guide that shapes all AI-generated text (tone, voice, formatting) so output matches your organization’s standards.

• Daily & weekly Daily editions synthesize entries from the last 24 hours into a hero story, themed sections, and source bibliography. Weekly editions summarize a week of daily output into broader narrative arcs.
• Scheduled generation Set editions to generate automatically on a daily or weekly schedule, or trigger manually.
• Custom branding Each edition can carry its own masthead, logo, favicon, accent color, and footer. Style guides provide editorial instructions for the AI writer’s tone and voice.
• Public URLs & sharing Each edition has a public URL at /edition/{slug} with a browsable archive index. Create expirable share links (7, 30, or 60 days) for external distribution.
• API access JSON and HTML export endpoints for integration with email newsletters, websites, or other systems.

• Research & analysis Analysts search across the organization’s entire corpus, combine internal documents with web sources, and share synthesized findings with their team. The AI grounds every answer in actual source material: no hallucination, no guesswork.
• Team knowledge base Groups organize teams and content. An analyst summarizes a report and shares it to the research feed. A writer pulls from the archive and sends findings to colleagues. Shared knowledge builds organically through AI-mediated collaboration.
• Content aggregation and monitoring Crawl and normalize hundreds of sources (industry publications, competitor blogs, wire services, regulatory feeds) into a single searchable interface. Automated digest delivers briefings from the sources and topics that matter to each user.

• CMS integration & replication Pull content from WordPress and other CMS platforms with incremental sync. Pluggable connector API for custom sources. Bulk importers for migrations. Everything lands in the same searchable index.
• Headless search API Lucene-quality full-text search as a service. Boolean queries, faceted filtering, configurable relevance, time-range constraints, and custom ranking signals, via GraphQL, REST, and MCP endpoints.