← all posts

How your AI connects to Darash, Junifye and Lexifye — MCP and DCR, explained plainly

Darash· ·4 min read ·Publifye AS ·publifye darash junifye lexifye

People assume that connecting an AI assistant to a professional tool means an afternoon of API keys, developer dashboards and configuration files. For Darash, Junifye and Lexifye, the honest description of setup is: paste one URL, sign in, done. This post explains why that is true and what is actually happening underneath — because the mechanism is worth understanding before you buy, not discovering after.

The one-URL path: Claude and Dynamic Client Registration

Every Publifye service speaks the Model Context Protocol (MCP) — the open standard that lets an AI assistant call external tools directly. In Claude, you add a custom connector, give it a name, and paste the service's MCP URL (Darash's, for example, is shown on its connector page). Then you sign in. That is the whole procedure.

What happens underneath is the part worth knowing. When Claude first talks to the server, it uses Dynamic Client Registration (DCR): your assistant registers itself as an OAuth client, on the spot. That is why there is no API key to copy, no client secret to guard, and nothing to install — the credential exchange that used to be a developer's chore happens machine-to-machine, in seconds, and the sign-in you see is an ordinary account login. (If a client ever asks for a Client ID, it is printed on the connector page — it is public, and the secret field stays blank, because there is no secret.)

Two properties fall out of this that people don't expect:

  • Set it up once, have it everywhere. Add the connector in Claude on the web and it is already there in the desktop app and on your phone — voice mode included. The engine answers in milliseconds, so a spoken conversation about a Hebrew word never stalls waiting on it.
  • Your account is the connection. Sign-in runs through the same Publifye account you got when you bought access — one identity across the store, the services, and the connector. There is no second registration.

Every other door in

  • Claude Desktop takes a four-line entry in claude_desktop_config.json — the exact snippet is on each connector page, copy-paste ready.
  • ChatGPT supports MCP connectors on the web, per OpenAI's own documentation — though not in voice mode, so we say exactly that rather than more.
  • GitHub Copilot speaks MCP too — point VS Code's agent mode at the same URL and it connects the same way, on free and paid plans alike (the free tier allows one connector, so make it count). A breeze, in our own daily use.
  • Cursor, Claude Code, and your own agents connect with a personal API key from your account — the classic route, still one URL and one key.
  • The cheapest route of all is the DeepSeek Harness: a free, open-source workspace you run on your own machine, paying DeepSeek directly for the model — cents per conversation — with no monthly AI fee stacked on top of your Publifye subscription. It is new and fast-moving (DeepSeek themselves call it a preview), and the connector pages walk through the whole install, Node.js version and all.

The side effect nobody advertises: less hallucination

One more thing, offered as our own experience rather than a benchmark: an assistant connected to Darash hallucinates less about Scripture, because it stops needing to. When the Hebrew, the parsing and the lexicon entry arrive as tool results with citations attached, the model quotes what it was handed instead of reconstructing from memory — and you can check every claim against the text, because the reference is right there. Grounding beats confidence. It is the same reason we made the engine hand over raw materials rather than interpretations in the first place.

Why this matters more than it looks

The three services share one authentication backbone, which is why they behave identically: buy Junifye, or Darash, or the Complete Trio on one key, and the account that purchase creates is the account your assistant signs in with. The connector works the minute the purchase completes. No provisioning email, no "allow 24 hours."

And because MCP is an open standard rather than our invention, you are never locked to one assistant. The same Darash that answers in Claude's voice mode on a walk answers in Cursor while you write, and in your own agent if you build one. The tools belong to your workflow, not the other way round — which is the whole reason we built them this way.

Each service's connector page has the URL, the snippets and the walkthroughs: Darash · Junifye · Lexifye. If you want to see what the tools actually do before connecting anything, start with how Darash compares to the software you may already know.