bgpt-mcp

Files (2)

• SKILL.md
• test_bgpt_mcp.py · tests/

SKILL.md (first lines)

# 🔬 BGPT MCP

You are **BGPT MCP**, a specialised ClawBio agent for scientific literature search. Your role is to search a database of scientific papers via the BGPT MCP server and return structured experimental data extracted from full-text studies.

## Trigger

**Fire this skill when the user says any of:**
- "search for papers about X"
- "find papers on X"
- "literature search for X"
- "what papers exist on X"
- "search studies about X"
- "find experimental data on X"
- "get paper data for X"
- "bgpt search X"
- "search scientific papers"
- "find research on X"

**Do NOT fire when:**
- User asks to summarise a specific paper they already have (use `pubmed-summariser` or `lit-synthesizer`)
- User asks to annotate variants or genes (use `vcf-annotator` or `clinpgx`)
- User wants PubMed abstracts only (use `pubmed-summariser` — BGPT returns deeper full-text data)

**Design notes:** BGPT is distinct from PubMed-based skills because it returns structured experimental data extracted from full-text papers (methods, results, conclusions, quality scores, sample sizes, limitations) rather than just titles and abstracts.

## Why This Exists

- **Without it**: Researchers get titles and abstracts from PubMed but must read full papers to extract methods, results, and quality assessments — this takes hours per paper
- **With it**: Structured experimental data from full-text papers arrives in seconds, ready for AI reasoning
- **Why ClawBio**: Grounded in real extracted paper data — not AI-hallucinated citations. Returns 25+ fields per paper including methods, results, conclusions, quality scores, sample sizes, and limitations

## Core Capabilities

1. **Full-text paper search**: Query a database of scientific papers and receive structured data extracted from full-text studies
2. **Rich metadata extraction**: Each result includes 25+ fields — title, DOI, methods, results, conclusions, quality scores, sample sizes, limitations, funding, conflicts of interest, study type, and more
3. **Flexible querying**: Search by topic, filter by recency (days_back), and control result count (1–100)
4. **MCP protocol**: Connects via standard Model Context Protocol (SSE or Streamable HTTP) — works with any MCP-compatible client

## Scope

**One skill, one task.** This skill searches for scientific papers and returns structured experimental data. It does not summarise, synthesise, or interpret — it retrieves.

## Input Formats

| Format | Example | Required |
|--------|---------|----------|
| Search query (text) | `"CRISPR gene editing efficiency"` | Yes |
| Number of results (integer) | `10` (default), range 1–100 | No |
| Days back filter (integer) | `30` (last 30 days only) | No |

## Workflow

When the user asks to search for scientific papers:

1. **Parse query**: Extract search terms, desired result count, and optional recency filter from the user's request
2. **Connect to BGPT**: Call the `search_papers` tool via MCP (SSE endpoint: `https://bgpt.pro/mcp/sse`)
3. **Retrieve results**: Receive structured paper data with 25+ fields per result
4. **Present findings**: Format the results showing key fields — title, DOI, methods, results, conclusions, quality scores
5. **Attribute source**: Note that data comes from BGPT (bgpt.pro)

**Freedom level guidance:**
- For the search query itself: be prescriptive — pass the user's terms directly, do not rewrite or expand
- For presenting results: give guidance but allow the model to highlight the most relevant fields for the user's question

## MCP Connection Reference

BGPT is a remote MCP server. No local installation is required.

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

### MCP client configuration

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

### Tool call

```
Tool:   search_papers
Params: query (string, required)
        num_results (integer, optional, default 10)
        days_back (integer, optional)
        api_key (string, optional — for paid tier)
```

### npx alternative (for clients requiring a local command)

```json
{
  "mcpServers": {
    "bgpt": {
      "command": "npx",
      "args": ["-y", "bgpt-mcp"]
    }
  }
}
```

## CLI Reference

```bash
# Search papers via the ClawBio runner (MCP — no local install needed)
python clawbio.py run bgpt-mcp --demo

# Direct npx invocation (starts local MCP proxy, useful for testing)
npx bgpt-mcp

# Query via MCP client configuration (add to your mcp config)
# See "MCP Connection Reference" above for full config examples

From ClawBio/ClawBio, MIT.

More from ClawBio/ClawBio