OpenMyst is an anchored-memory retrieval tool. Ask it anything; it answers only with verbatim excerpts from real web pages—each pinned to its source and exact line. Connect it to Claude as an MCP tool, or run the standalone research app.
Free plan, no credit card · open source · works in Claude desktop, web, and code
The model is already good enough. The missing piece is making sure every sourced claim is real. That's the whole project.
A model that invents a source—a plausible author, a real-sounding journal, a page number that doesn't exist—is worse than one that says nothing. Most of the harm from LLMs in serious work traces back to confident, unverifiable references. That's the single failure mode this project exists to remove.
OpenMyst never paraphrases from memory. Every fact it returns is a verbatim substring of a real page it fetched, located deterministically down to the exact line. If a claim can't be pinned to text that provably exists, it isn't returned. No anchor, no claim—the load-bearing rule of the whole system.
This isn't a chatbot trying to replace your tools. It's a small, sharp retrieval layer that bolts onto the model you already use. Connect it to Claude as an MCP tool and it does the expensive reading off to the side, handing back only cited evidence. The standalone app is here too—but the point is the grounding, not the wrapper.
Adding the connector gives Claude a research tool. When you ask it to look something up, it hands the job to OpenMyst, which searches the web, reads the sources, and returns verbatim, line-cited evidence—so Claude can reference every claim instead of guessing.
https://mcp.openmyst.ai/mcp“Find sourced evidence on whether Pareto fronts are used in neural architecture search.” Claude decides to call the research tool.
Numbered verbatim excerpts, each with its source and exact line, plus a ready-to-paste Sources section—so Claude cites inline instead of hallucinating.
OpenMyst is a standard remote MCP server, so the same connector URL drops into any MCP-capable client. Here's the two-minute setup for the popular ones.
https://mcp.openmyst.ai/mcpAvailable on ChatGPT Pro, Team, Enterprise & Edu via Developer Mode connectors.
The expensive reading happens server-side with cheap models, off your agent's context. What comes back is small, relevant, and provably real.
A fast model turns your natural-language request into a handful of focused web queries—disambiguating the topic so it searches for the thing you actually mean.
It searches the web (Brave, with a free fallback) and reads each page with its own cheap models—off your context budget. PDFs, papers, articles, all the way through.
A cheap extractor pulls typed snippets—definitions, claims, statistics, findings—each a verbatim excerpt. Pure code then locates each one in the source and records its exact line. Paraphrases and ambiguous matches are dropped.
A final pass keeps just the anchors that bear on your request and writes a short overview. You get a tight, cited set—never the whole corpus, never a wall of scraped text.
No vector database, no embeddings. Anchors are located by exact string match and stored with byte offsets—retrieval is a deterministic read, not a similarity guess.
The same anchored-memory engine ships as a standalone desktop app—a multi-agent research collaborator that plans, gathers the literature, and drafts a fully cited piece. Same grounding rule, more of the workflow. Entirely optional; the connector is the heart of the project.
Explorer, Skeptic, Steelman, Architect, Adversary, Audience and others read your work each round, run autonomous searches when the literature has gaps, and surface the 2–3 sharpest questions only you can answer—with role attribution and provenance back to your own sources.
Every source lives in .myst/wiki/ as plain markdown and JSON, rendered as a force-directed graph of inferred source-to-source links. The panel searches the web for what's missing while it deliberates, and new sources land anchor-indexed and ready to cite. No embeddings, no vector DB—just an inverted index you can grep.
The drafter can only cite anchors that exist on disk. Every inline citation is a link whose fragment points to the exact line of the exact source; hover it and you see the source's actual words. Bibliographic metadata is extracted at ingest, so citations are real Harvard format from the first draft.
macOS · Windows · Linux · free, same account
Free forever for the connector and the app. Five dollars a month if you go heavy.
Plenty to use the Claude connector daily.
For heavy daily use. Unlimited research.
Cancel anytime. 7-day money-back guarantee.
Connect the tool to Claude in 90 seconds—free, open source, no credit card.