Research and data workflows with Muse

Muse handles research and data work in two modes. The personal agent is good at long, tedious collection jobs across your own accounts and public sources, like searching every state's unclaimed property database or merging 400 Google Docs. Muse Spark and Muse Glimmer, used directly, suit document extraction and analysis, with Spark's 1M-token context for big inputs and Glimmer for data that should stay local.

What research tasks has the Muse agent done?

From the catalog, with results as reported by the authors:

Build	Task
Unclaimed property search	Searched every state's database by name and past cities; found $88.92
400 Google Docs into one PDF	Combined 400 docs into a 400-page PDF plus HTML, Word and TXT
Census of a 77k story archive	Paged through an Instagram archive with resumable cursor state
Spotify playlists by decade	Re-sorted 18 playlists of about 200 songs each
Receipt audit	Six months of email receipts, one $1,200/year waste found
Resurfacing forgotten records	Found an 8-year-old ticket resale profit in about a minute

Two design lessons repeat. Long jobs need checkpoints: the 77k census saves cursor state every 10 pages so it can resume. And scale still breaks things: the Spotify author reports a similar job on 40,000 Google Photos failed.

Can Muse Spark analyze data and find sources?

Yes. Alexandr Wang had Muse Spark find open data and analyze a century of global GDP share. Fei Xia split an object-counting task across parallel subagents and reports much better results than a single pass, a useful trick for any count or tally.

Tests of its limits, as reported:

• Deedy's visual grounding tests found counting and bounding boxes "far from perfect."
• Simon Willison dug into the Code Interpreter and visual grounding tools in Meta AI's chat.
• Evergreen Capital found it comparable to Opus 4.6 on web data search and PDF parsing, though not quite frontier.
• A calorie estimation benchmark had Spark 1.3 lead with 48% of estimates within 20% error.

How much context does Muse Spark handle for research?

Spark 1.3 accepts 1,048,576 tokens, with text, image and video/document input (Meta docs). Meta reports MRCR v2 long-context retrieval of 98.5 at 256K–512K and 98.1 at 512K–1M (MarkTechPost). Those are Meta's numbers.

Web search grounding costs $2.50 per 1,000 queries on the API, and cached input is $0.15 per million tokens on standard (pricing). If you query the same corpus repeatedly, caching is where the savings are.

Can I do private research locally with Muse Glimmer?

Glimmer runs offline, handles text and images, and Meta reports 74.6 on DeepSearch QA (model card). Catalog examples:

• dealscan reads music industry contracts, extracts payment terms and flags risky clauses as markdown and JSON.
• Chaty is an offline desktop app with a RAG knowledge base and deep research on Glimmer.
• Glimmer HD Vision tiles 4K images for OCR and layout extraction.
• Cobus Greyling's agent lab lets you explore agent loops and memory before downloading weights.

Glimmer's knowledge cutoff is January 4, 2026, so pair it with a search tool for anything recent.

Can Muse work with health and medical data?

Two launch-era connectors bring personal health data in. Function Health links lab results and clinician-reviewed summaries. HealthEx brings medical records so Muse can summarize visits, draft questions for a doctor and set refill reminders. Treat anything it produces as notes to bring to a clinician.

Where do researchers go wrong with Muse?

• Trusting the summary. Several builders, like the delegate-to-muse author, explicitly check outputs rather than trusting the report.
• One pass on big jobs. Fan out, as Fei Xia did, and checkpoint, as the story census did.
• Assuming benchmark efficiency. DataCamp's tests found Spark 1.3's token savings varied by task, with a net 12% cost increase across three (DataCamp).
• Ignoring quota. Muse quotas are in tokens with no published task conversion (Layer3 Labs).

A research workflow template for Muse

The builds above share a shape you can reuse for your own project:

1. Scope the sources. Name the accounts, sites or folders Muse should read, and set connectors to read-only.
2. Define the output first. A table with fixed columns, a JSON schema or a one-page brief. Spark supports structured output on the API, and a fixed format makes checking easier on the agent too.
3. Checkpoint long runs. Save progress to a file every N items so the job can resume after a quota limit or error.
4. Fan out when counting. Split tallies across subagents and compare, as in the object-counting test.
5. Spot-check. Pick ten rows at random and verify them by hand before you trust the rest.
6. Keep sources. Ask for a link or file reference on every claim, so you can audit later.

Useful reference lists

• Awesome Meta Muse use cases: 156 source-linked examples in 16 categories.
• Meta Muse connectors, full list.
• Barbaros MCP: a read-only connector serving a quote-sourced dataset on 28 AI apps' privacy answers.

Browse research and data. Related: Muse for trading and markets and email triage with Muse.

Frequently asked questions

Is Muse good for deep research?

It does well on long collection jobs across accounts and public sites, and Spark 1.3 has a 1M-token context. Independent testers found it strong on web search and PDF parsing but not quite frontier.

Can Muse analyze spreadsheets and PDFs?

Yes. Muse Spark accepts document input on the API, and builders have used Muse to build Excel order forms, merge 400 Google Docs and parse PDFs.

Can I run Muse research workflows offline?

Use Muse Glimmer 30B, which runs locally under Apache 2.0. Its knowledge cutoff is January 2026, so add a search tool for recent material.

How do I stop long Muse research jobs from failing?

Split work into parallel subagents and save progress at checkpoints so jobs can resume. The Instagram story census saved cursor state every 10 pages.