Documentation

claude mcp add -s user --transport http science https://api.scientistworkbench.com/mcp \
  --header "Authorization: Bearer wb_live_…"

Then /mcp inside Claude Code lists the server; tools appear as mcp__science__science_submit etc. Set MCP_TIMEOUT generously or loop on science_wait.

{
  "mcpServers": {
    "science": {
      "url": "https://api.scientistworkbench.com/mcp",
      "headers": { "Authorization": "Bearer wb_live_…" }
    }
  }
}

Cursor: .cursor/mcp.json. Claude Desktop and other clients that only speak stdio can run uvx --from scientist-workbench workbench mcp with WB_URL=https://api.scientistworkbench.com and WB_API_KEY in the environment.

from claude_agent_sdk import ClaudeAgentOptions, query

opts = ClaudeAgentOptions(
    mcp_servers={"science": {"type": "http", "url": "https://api.scientistworkbench.com/mcp",
                             "headers": {"Authorization": "Bearer wb_live_…"}}},
    allowed_tools=["mcp__science__science_submit", "mcp__science__science_wait",
                   "mcp__science__science_result", "mcp__science__science_artifact"],
)
async for msg in query(prompt="QC the attached CRO IC50 report …", options=opts):
    print(msg)

curl -s https://api.scientistworkbench.com/jobs -H "Authorization: Bearer wb_live_…" \
  -F 'spec={"prompt":"QC this CRO IC50 report and say which values can be used as reported","company":"Acme"}' \
  -F files=@plate1.xlsx
# → {"job_id":"sci-20260914-…","status":"queued"}
curl -s https://api.scientistworkbench.com/jobs/sci-20260914-…/result -H "Authorization: Bearer wb_live_…"