Using SPARKIT from Claude Desktop, Cursor, and Claude Code

The most common workflow we hear from researchers is: open Claude or Cursor, sketch out the question, leave the LLM, go run a literature search, paste the results back in, ask the LLM to synthesize. Three context switches and a lot of copy-paste for what is mostly the same task: "go look up the literature on X."

sparkit-mcp collapses that loop. SPARKIT is now a tool the LLM client can call directly, through Anthropic's Model Context Protocol. One config block, restart your client, and the model can hand off literature questions to SPARKIT mid-conversation and get back a cited Markdown report.

Install

uv tool install sparkit-mcp

(or pip install sparkit-mcp)

That puts a sparkit-mcp console script on your PATH. Then point your MCP client at it.

Claude Desktop

Edit ~/Library/Application Support/Claude/claude_desktop_config.json on macOS (or %APPDATA%\Claude\claude_desktop_config.json on Windows):

{
  "mcpServers": {
    "sparkit": {
      "command": "sparkit-mcp",
      "env": { "SPARKIT_API_KEY": "sk_sparkit_..." }
    }
  }
}

Restart Claude Desktop. The sparkit tool appears in the tools menu next to the chat input.

Cursor

Edit ~/.cursor/mcp.json (or .cursor/mcp.json in your project) with the same block. Reload the window.

Claude Code

claude mcp add sparkit -e SPARKIT_API_KEY=sk_sparkit_... -- sparkit-mcp

Get an API key if you don't already have one.

Two tools

The server exposes two:

• research(question, ...) submits a scientific question. SPARKIT searches the literature, reads the relevant papers, and returns a cited Markdown report. The tool blocks until the job finishes (default 4 minutes) and surfaces the report inline.
• get_job_status(job_id) fetches a job by id. Used when research returned before the job finished, or to revisit a past report.

The model picks which one to call. In practice, research is what gets used 95% of the time.

What changes

A few workflows that get faster:

Reading a paper, want a fact you do not trust the LLM with by itself. "Look up the actual citation count and replication status for the original BAFF receptor knockout study in mice." The model hands off to SPARKIT, gets a cited answer back, and continues without you copy-pasting anything.

Coding agent that needs literature context. A Cursor agent debugging a bioinformatics pipeline can ask "what does the latest literature say about X normalization for bulk RNA-seq?" mid-task, instead of stopping and asking the human to go look it up.

Writing a grant or paper. Open a draft in Claude Desktop, ask "is the consensus on Y still that Z, or has the field updated since 2024?" The model calls SPARKIT, gets a report with sources, and helps revise.

The unifying property: the LLM stays in the driver's seat, and SPARKIT becomes the literature substrate it pulls from.

Latency and cost

SPARKIT jobs typically complete in 60-180 seconds. The tool blocks during that wait; Claude Desktop and Cursor both show a "thinking..." state while SPARKIT works. For deeper questions, more sources, more iterations, the wait can stretch to 4-5 minutes. The tool defaults to a 4-minute hard timeout; if a job exceeds it, the tool returns the job_id and the model can come back via get_job_status later.

Each research call consumes one query against your SPARKIT plan. Plus is 15/mo, Pro 35/mo, Max 120/mo, and overage is $5/query (Plus and Pro) or $4/query (Max). The MCP path doesn't change pricing, it's the same /v1/research call your SDK was already making.

What's missing

This is a v0.1 release. Things on the near roadmap:

• Streaming progress. MCP supports progress notifications; we are not sending them yet, so the client UI just shows "thinking" until the job completes. Adding granular progress (searching → reading → drafting) is straightforward and lands soon.
• Resource exposure. The MCP resources capability lets a server hand the client structured documents. We could expose past reports as resources so the model can pull them back into context. Not in v0.1.
• More controls. response_format and include_citations are exposed; deeper controls (max sources, target length) aren't yet.

Source

Open-source under MIT: github.com/SPARKIT-science/sparkit-mcp. Issues and PRs welcome. If you wire up a different MCP client (Zed, Continue.dev, etc.), tell us at info@sparkit.science and we'll add it to the README.