Evidence infrastructure for AI

The Evidence Layer for Scientific AI

Give your AI evidence you can verify before you trust its answer.

6,073
new studies added this week
Keep your AI closer to the latest evidence.

Build scientific AI that can show its work.

Search by topic or DOI. Retrieve claims linked to experiments, exact results, scope, limitations, falsification criteria, and source provenance.

When an answer matters, you can trace it back to what each paper actually showed.

50 free results · Then $0.02/result

Check the answer See the experiments and exact results behind each claim.
Trace it to the source Follow evidence to source sections, figures, tables, and text.
See where it stops Keep scope, biases, extraction status, and uncertainty visible.

Built for scientific model training · literature review · RAG · research agents


Quick setup

Get Started in 30 Seconds

Pick your tool, copy the config, and start searching.

Add to ~/.claude/claude_desktop_config.json:

{
  "mcpServers": {
    "bgpt": {
      "url": "https://bgpt.pro/mcp/sse"
    }
  }
}

Add to .cursor/mcp.json in your project root:

{
  "mcpServers": {
    "bgpt": {
      "url": "https://bgpt.pro/mcp/sse"
    }
  }
}

Run in your terminal:

claude mcp add bgpt --transport sse https://bgpt.pro/mcp/sse

Traditional HTTP API. No key needed for the first 50 results.

import requests

def search_bgpt(query, num_results=16, days_back=None, api_key=None, output_format="evidence"):
    payload = {
        "query": query,
        "num_results": num_results,
        "output_format": output_format,
    }
    if days_back is not None:
        payload["days_back"] = days_back
    if api_key:
        payload["api_key"] = api_key

    response = requests.post(
        "https://bgpt.pro/api/mcp-search",
        json=payload,
        timeout=60,
    )
    response.raise_for_status()
    return response.json()["results"]

papers = search_bgpt("CRISPR delivery neurons", num_results=5)
paper = papers[0]
print(paper["central_claim"])
if paper["evidence"]["evidence_units"]:
    unit = paper["evidence"]["evidence_units"][0]
    print(unit["claim"])
    print(unit.get("result", {}).get("reported_statistics", []))
    print(unit["provenance"])

Endpoints for any MCP-compatible client:

Endpoints:
  SSE:              https://bgpt.pro/mcp/sse
  Streamable HTTP:  https://bgpt.pro/mcp/stream

Tools:

  search_papers
    query        (required)  Search terms
    num_results  (optional)  1-100, default 16
    days_back    (optional)  Filter to recent papers
    output_format(optional)  evidence, full, or legacy

  lookup_paper
    doi          (required)  DOI of the paper
    output_format(optional)  evidence, full, or legacy

From question to evidence

How It Works

  1. Add the config above to your AI tool
  2. Ask your AI to search, e.g. “Search BGPT for CRISPR delivery methods in neurons” or look up a specific paper by DOI
  3. Get evidence units linking claims to experiments, results, scope, and provenance

Your first 50 results are free. After that, add billing at $0.02/result.


Simple metered access

Pricing

One result is one structured paper record.

Free Tier

$0

  • ✓ 50 free results
  • ✓ Evidence ready for training and RAG
  • ✓ No API key needed

Pay-as-you-go

$0.02 / result

  • ✓ Unlimited searches
  • ✓ Evidence ready for training and RAG
  • ✓ Pay only for returned results
Keep Searching for $0.02/Result →

Inspect the record

What BGPT Returns

Use output_format="evidence" for compact records or "full" to add legacy metadata.

Search responses place records in results. DOI lookup places the record in result and includes found. Evidence and full output omit papers without extracted evidence.

Identification
doi: DOI identifier
title: Paper title
publication_date
publication_name: Journal
Claim-Level Evidence
central_claim
evidence.schema_version and extraction_status
evidence.document_type
evidence.evidence_units: claims linked to experiments
claim_type: causal, correlational, mechanistic, descriptive, diagnostic, prognostic, efficacy, or safety
evidence_status: supported, mixed, contradicted, or not_assessable
evidence_basis: direct, indirect, or synthesized
relation_to_central_claim: supports, qualifies, challenges, or independent
experiment: design, independent units, groups, outcomes, and methods
reported_statistics: exact reported value text, comparison, test, and analysis model
applicability.demonstrated_in: demonstrated scope
provenance[]: section, figure, table, exact source text
evidence.record_provenance: source hash, extraction time, truncation status
Scientific Scrutiny
counterevidence_or_ambiguity[]: reported or inferred qualifiers
limitations[]: reported or inferred limits
evidence_units[].falsification_criteria[]: concrete weakening tests with basis
how_to_falsify: concise paper-level criterion
implicit_assumptions: useful tacit knowledge the paper leans on but never states
limitations_and_biases: concise paper-level synthesis
Paper-Level Fields in Full Output
sample_size_and_population_characteristics
methods_and_experimental_techniques
results_and_conclusions
experimental_models
software_and_tools_used
code_and_data_links
Additional Fields in Full Output
keywords
lab_name
funding_json
citations_json
reference_count
legacy metadata: retained for existing clients

Coverage reflects the available full text and what the paper reports. Empty optional fields are omitted from compact evidence output. V4 records contain at most 5 evidence units and 2 provenance passages per unit; each passage is at most 320 characters. Older records retain their explicit schema version. extraction_status and any normalization_warnings expose incomplete or repaired records.


Details and support

Frequently Asked Questions

How do I start searching for free?

Just add the config above to your AI tool and start asking questions. No sign-up, no API key needed. You get 50 free results automatically.

What happens when I use up my 50 free results?

How do I start pay-as-you-go?

How do I use my API key?

How does billing work exactly?

What does days_back do?

Where do I find my API key if I lost it?

Can I cancel anytime?

Need help?


Stop taking scientific answers on faith.

Give your AI evidence you can inspect. The first 50 records are free.

Copy your config