SPARKIT | Scientific research agent API

# SPARKIT > A scientific research agent API. SPARKIT searches the web and scientific literature, reads and compares evidence, runs calculations or code when needed, and returns an inspectable cited report. SPARKIT is designed for developers calling research from an agent, backend, or scientific workflow. A human-facing research interface is also available in the SPARKIT dashboard. ## Core capabilities - Search web sources and scientific literature - Read papers, PDFs, and web pages - Compare evidence and identify caveats - Write and execute code for calculations and analysis - Return Markdown reports with inline citations and structured sources when literature evidence is relevant - Run asynchronously through polling or webhooks ## Public benchmark results Measured June 2026. Exact systems and modes are named below; SPARKIT does not generalize these results to other product modes or model versions. - **HLE-Gold:** 149 questions; biology / medicine + chemistry - SPARKIT: 54.4% (SPARKIT research-agent run) - GPT-5.6-Sol: 39.0% (Direct model call) - Claude Opus 4.8: 34.9% (Direct model call) - Methods and limitations: https://sparkit.science/benchmarks/hle-gold - **GAIA:** 127 questions - SPARKIT: 73.2% (SPARKIT research-agent run) - Exa: 58.2% (Search API result) - Brave: 57.0% (Search API result) - Methods and limitations: https://sparkit.science/benchmarks/gaia Canonical benchmark index: https://sparkit.science/benchmarks ## Interfaces - API documentation: https://sparkit.science/docs - Python SDK: https://pypi.org/project/sparkit-science/ - MCP integration: https://sparkit.science/docs/integrations - Dashboard: https://app.sparkit.science/research ## Commercial information - Current plans and terms: https://sparkit.science/pricing - Custom engagements: https://sparkit.science/services ## Important limitations SPARKIT is an LLM-driven system and can be wrong. Citations may be misattributed, sources may be summarized inaccurately, and co