API and MCP

Core in your own assistant.

Connect Claude, ChatGPT, Cursor, Codex or Gemini CLI to your Core. They read your library and data as you, with your access, and every call is logged.

Claude answering a question with Core over MCP: it calls search_documents and graph_lookup on your Core, gets numbered passages with document and passage ids, and cites them in its answer.

How a call reaches your Core.

  1. Your assistant

    Claude, ChatGPT, Cursor, Codex or Gemini CLI, under your own account.

    MCP over HTTP

  2. Your token

    One per person, from Core. It acts as you, with exactly your access.

    cmt_…

  3. The relay

    Claude.ai and ChatGPT call from the internet. The relay reaches your Core without opening your network.

    TLS on your machine

  4. Your Core

    Checks the token and your access, then runs the tool. 60 calls a minute, 4 at once.

    per token

  5. Back to you

    Numbered passages and sourced links, and a log line: who, which tool, how long.

    audit log

What your assistant can use.

ToolWhat it doesTry asking
search_documentsSearch your library by meaning and keyword. Returns numbered passages to cite.“What did management say about HBM supply in Q2?”
read_documentRead the passages around one you already have: a table’s continuation, the rest of a list.“Read the rest of that guidance table.”
graph_lookupFind companies, funds, people, products and deals in your map, each link with its quote.“Who supplies NVIDIA’s GB300?”
fund_snapshotneeds book accessThe book: positions, exposure and movers.“What moved our book overnight?”
market_closeThe latest stored daily close for a ticker, as a numbered source.“Where did TSMC close yesterday?”
data_queryOne read-only SQL query over the datasets you loaded.“Sum Q3 fund flows by strategy.”
data_describeA dataset’s columns, sample rows and statistics.“What is in the fund flows table?”
data_pack_searchneeds licenceSearch a licensed Data Pack.“Find recent rounds in Korean fintech.”

Connect in a minute.

  • Terminal
    $ claude mcp add --transport http core https://your-core.example.com/mcp \
      --header "Authorization: Bearer cmt_…"

    Claude Code

    One command adds your Core as a connector.

  • Add custom connector

    Remote MCP server URLhttps://your-core.example.com/mcp/cmt_…

    CancelAdd

    Claude.ai and ChatGPT

    Paste one address as a custom connector. The token is in it, because these apps have no header field.

  • POST /chat
    curl -N https://your-core/api/chat \
      -H "Authorization: Bearer …" \
      -d '{"question": "…"}'
    
    event: step  {"kind":"retrieve"}
    event: delta {"text":"# NVIDIA…"}
    event: done  {"tokens_in":7180}

    REST

    Ask itself, over HTTP: the answer streams as events, the same ones Core’s own page reads.

  • tools/list
    POST /mcp
    {"jsonrpc": "2.0", "id": 1, "method": "tools/list"}
    
    { "name": "search_documents",
      "description": "Search the firm's documents.
        Returns numbered passages [n] to cite.",
      "inputSchema": {
        "required": ["query"],
        "properties": {
          "query": {…}, "company": {…},
          "period": {…}, "limit": {…} } } }

    Any MCP client

    What any MCP client sees: each tool with its description and input schema.

Ask in Core, or your own assistant.

  • Search, documents and the map

    Ask in Core
    Yes
    Your assistant over MCP
    Yes
  • Your book and licensed data, with access

    Ask in Core
    Yes
    Your assistant over MCP
    Yes
  • Acts as you, every call logged

    Ask in Core
    Yes
    Your assistant over MCP
    Yes
  • Where the reasoning happens

    Ask in Core
    Your model, on your machine if you choose
    Your assistant over MCP
    Your assistant’s provider; passages go to it
  • Your firm’s rules for an answer

    Ask in Core
    Yes
    Your assistant over MCP
    A short note on how to cite
  • One citation list for the whole answer

    Ask in Core
    Yes
    Your assistant over MCP
    Numbered per call
  • Every figure checked against its sources

    Ask in Core
    Yes
    Your assistant over MCP
    No
  • Your position in the company, shown

    Ask in Core
    Yes
    Your assistant over MCP
    No
  • Saved for review, in the answer log

    Ask in Core
    Yes
    Your assistant over MCP
    Call log only
  • Virtual analyst and licensed research

    Ask in Core
    Yes
    Your assistant over MCP
    No

Use Ask for work your firm will rely on. Use your own assistant to bring your library into the tools you already use.

In your hands.

One token per person
Each acts as that person, with exactly the access they have in Core.
60 calls a minute, 4 at once
Per token, so one script can’t crowd out the desk.
Refused when it should be
A revoked token, an inactive person or an unfinished sign-in step is refused, and recorded.
Rotate any time
A new token retires the old one at once.
Off with one setting
Admins can turn MCP off for the whole Core.
Your account, your assistant
Core never holds your Claude or ChatGPT sign-in.
The record each MCP call writes: who, which tool, whether it succeeded, how long it took, how much it returned, and the start of its arguments.

Questions.

What leaves our machine when someone uses this?

The passages and data the assistant asked for go to that person’s assistant, under their own account with Anthropic, OpenAI or another provider. The security page lists this with every other connection.

Can a person see more through MCP than in Core?

No. The token acts as that person, with the access they have in Core when the call is made. A deactivated user is refused.

Why use Ask in Core if we have Claude?

Your own assistant reads the same library, but it doesn’t carry your firm’s rules for an answer, check every figure against its sources, or save the answer for review. Use Ask for work the firm relies on, and MCP to bring the library into the tools you already use.

How is it priced?

API and MCP access is an add-on to your Core. See the pricing page.

Which assistants work?

Claude.ai, Claude Code, Claude Desktop, ChatGPT (developer mode), Codex CLI, Gemini CLI and Cursor. Anything that speaks MCP over HTTP should work.

Explore more of Core.

Bring Core to your tools.

Ask for API and MCP access with your Core.