oraclip-mcp — the entry product

An MCP server that injects a curated, regularly-updated, verbatim-cited operator brain (YC, a16z, Sequoia, Lenny's Podcast, Anthropic, Andrej Karpathy, AI Engineer, Latent Space, Dwarkesh, SaaStr — 135+ videos, ~6,000 chunks; nightly refresh target — see operator_brain_stats for the last build timestamp) into the tools you already use: Claude Code, Cursor, Cowork, Claude Desktop, ChatGPT desktop.

Why this exists (the wedge)

You're not going to switch away from Claude/Cursor — so don't. This rides inside them. The big models answer from general, stale, uncited knowledge. This tool returns the actual words of the best operators, with the speaker, the date, and the source YouTube clip, on a regularly-updated niche corpus (nightly refresh target) — the one thing they're structurally weak at. It augments the host model; it doesn't compete with it.

Tools

• ask_operators(query, top?, theme?) → verbatim, cited passages (channel · date · clip URL) for the host to ground + cite its answer.
• operator_brain_stats() → coverage + freshness.

Install

The default path is the hosted brain — nothing to run locally, no Python, no index. Two thin ways to reach it (A and B); a fully-local self-host option is in Advanced below.

Getting a key

There's a free shared key baked into the configs below — paste-and-go, no signup. It's a fair-use shared tier (a daily call quota). For a higher-quota key of your own (plus the priority-latency / freshness SLA), email [email protected] (self-serve signup lands with billing at S5). Pass the key as RETRIEVER_TOKEN (npx path) or as the Bearer token (remote-MCP path).

Free shared key: ok_free_2620edf301e0560401b4d002290c4b38642b2f8e

A. Hosted remote MCP (recommended — zero local setup, nothing to run)

For hosts that support a remote MCP endpoint, point them directly at the hosted /mcp endpoint with the key as a Bearer token — no npx, no local process:

URL:  https://oraclip.agni.work/mcp
Auth: Bearer ok_free_2620edf301e0560401b4d002290c4b38642b2f8e

B. npx thin client → hosted brain (for hosts that launch a local command)

For hosts (Cursor / Claude Desktop / Cowork config) that spawn a local stdio command, run the thin oraclip-mcp client and point its RETRIEVER_URL at the hosted retriever base (the client appends the retriever paths itself — do NOT put /mcp here; /mcp is the remote-MCP endpoint in path A, a different surface):

{
  "mcpServers": {
    "operator-brain": {
      "command": "npx",
      "args": ["-y", "oraclip-mcp"],
      "env": {
        "RETRIEVER_URL": "https://oraclip.agni.work",
        "RETRIEVER_TOKEN": "ok_free_2620edf301e0560401b4d002290c4b38642b2f8e"
      }
    }
  }
}

Equivalent one-liner:

RETRIEVER_URL=https://oraclip.agni.work RETRIEVER_TOKEN=ok_free_2620edf301e0560401b4d002290c4b38642b2f8e npx -y oraclip-mcp

Then ask your agent: "Use operator-brain: what do the best operators say about pricing an AI product?" — it returns cited clips and grounds the answer.

Free vs. paid

What's live today is the honest differentiator:

• Free: the shared key above — a fair-use daily call quota against the same hosted brain, no signup. Enough to evaluate and use it in real workflows.
• Paid: your own key with a higher (or removed) daily quota plus a priority-latency / freshness SLA (priority embed lane, guaranteed refresh-recency target). Email [email protected].

Both tiers query the same corpus today. Corpus-scoping and per-seat features land at S5 — so the free tier is not a reduced corpus subset (that gating isn't built yet); the live distinction is quota + (paid) latency/freshness SLA.

Env

• RETRIEVER_URL — hosted retriever base (e.g. https://oraclip.agni.work). If set, no local Python is used.
• RETRIEVER_TOKEN — bearer key for the hosted retriever. A free shared key is included above; for your own higher-quota key, email [email protected].
• OPERATOR_BRAIN_VAULT (Advanced / self-host only, default ai-content-brain)
• GBRAIN_PORT (Advanced / self-host only, default 8079)

Advanced / self-host (run the retriever on your own machine)

Most users should use the hosted path above. If you want to run the whole retriever locally (no hosted dependency), this repo can spawn a local Python retriever daemon.

Architecture: thin Node MCP (stdio) → local Python retriever daemon (gbrain_serve.py, holds the embedding model + vector matrix in memory; auto-started on first call). Host model = reasoning; this tool = cited retrieval. Retrieval ~0.3s.

Prereqs: cd agni-sage/mcp && npm install · Python fastembed+numpy set up. With no RETRIEVER_URL set, the retriever auto-starts on first call:

claude mcp add operator-brain -- node "/absolute/path/to/agni-sage/mcp/server.mjs"

Cloud / commercialize

Cloud deploy: a single stateless Node service on Vercel scale-to-zero (~$0 idle; query embedding offloaded to Together serverless). Hosted at https://oraclip.agni.work. Deploy details live in deploy/vercel/DEPLOY.md.

Monetization path

This is the entry product: distribution = the host tools' users; the moat = the curated, cited, regularly-updated corpus they can't reproduce. Meter by usage/seat; every call is also flywheel data. The free public "cited clips" page is the funnel; the Distillery (mint cited models from any corpus) is the B2B expansion.