Use Hjarni with Meta Muse

Where Muse stands on MCP

That surprises people, because Muse Spark, the model underneath, is explicitly trained for it: Meta's launch post says it “zero-shot generalizes to new native tools, MCP servers, and custom skills.” The model can speak MCP. The consumer app gives it nothing to speak to, and the Meta Model API documents a web_search tool and ordinary function calling with no MCP tool type, so even a developer wiring Muse Spark up runs the MCP client themselves.

So the honest summary is two different answers to one question, and the one that matters depends on which Muse you mean:

If Muse ever grows a field that takes a server URL, Hjarni works in it that day with nothing to build: https://hjarni.com/mcp is the same endpoint Claude and ChatGPT already use. This page exists because that field does not exist yet.

Connect Muse to Hjarni

The API contract Muse needs

Everything below is what a custom connector has to know. The Markdown twin of this page is the same content without the markup, which is the URL to hand any agent. The full REST API reference documents every endpoint, including the ones Muse will not need.

Base URL

https://hjarni.com/api/v1

Auth

Authorization: Bearer YOUR_API_TOKEN

Content type

application/json

Pagination

page and per_page query params, 25 per page by default and 100 at most. The total comes back as an X-Total-Count header, with X-Page and X-Per-Page beside it.

Rate limit

100 requests per minute, per account rather than per token, so this connector shares one budget with every other token on the account. Over it returns 429 with Retry-After.

Response shape

No envelope on the way out: collections are bare JSON arrays, a single note or folder a bare object. Writes take the {"note": {...}} wrapper, reads do not return one.

Note format

Markdown body, plus a short summary the agent should always write.

The six calls that cover ordinary use

Worked examples

Run these in a terminal first, so a later failure has one fewer cause. A working curl rules out the token and the endpoint; it does not by itself prove the remaining fault is Muse's, since a connector can also be sending a shape the API refuses.

Confirm the token

curl https://hjarni.com/api/v1/me \
  -H "Authorization: Bearer YOUR_API_TOKEN"

Search, titles and summaries only

curl "https://hjarni.com/api/v1/search?q=quarterly+planning&search_scope=personal&exclude_body=true" \
  -H "Authorization: Bearer YOUR_API_TOKEN"

Create a note

curl -X POST https://hjarni.com/api/v1/notes \
  -H "Authorization: Bearer YOUR_API_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "note": {
      "title": "Kitchen renovation quotes",
      "body": "- Vanhoof: EUR 14,200\n- De Smet: EUR 12,800",
      "summary": "Two quotes for the kitchen, with prices and lead times.",
      "tag_list": "house, renovation",
      "container_id": 12
    }
  }'

That container_id is an example. Take a real one from GET /containers: an id that is not yours answers 404, and leaving it out files the note at the root. A request carrying no note object at all answers 400.

Two shapes, and a connector that reads only one reports the other as a blank failure: a single error string, or an errors array when a note fails validation. Most errors are JSON, but not all of them are guaranteed to be, so send Accept: application/json and treat a non-JSON error body as possible rather than impossible.

What to ask Muse once it is connected

On an ordinary account-wide token, Muse reads and writes the same personal notes Claude and ChatGPT see. One asymmetry to know: search spans your team notes while fetching and writing a note stays personal, so a team hit in search results is not necessarily one this token can open. The useful prompts are the ones that lean on Muse's own connectors at the same time, since that is what it has and your other assistants do not:

• Across Gmail and your notes. “Find the contractor emails from last month and add anything new to my kitchen renovation note in Hjarni.”
• Before a meeting. “My calendar says I am seeing Anna at 3. Search Hjarni for what we agreed last time.”
• Capture on the phone. “Save this as a Hjarni note in my Ideas folder, with a summary.”
• Keep one note true. “Update my ‘working preferences’ note rather than making a new one.”

Tell Muse once to always write a two or three sentence summary on a new note. Summaries are indexed for search and are what a result list carries in place of a full body, so a connector that skips them leaves every other assistant less to go on when it picks which note to open.

What this route costs you

A custom connector is a worse connection than MCP, not an equal one. Worth knowing which parts you are giving up, because most of them are invisible until something goes wrong:

If Muse is your only assistant, this is a reasonable connection and the notes stay yours either way. If you also use Claude or ChatGPT, connect those over MCP and treat Muse as a second reader of one knowledge base. MCP vs API is the longer version of this argument.