Skip to content

CONNECT_YOUR_TOOLS.md

Latest commit

 

History

History
81 lines (62 loc) · 3.61 KB

CONNECT_YOUR_TOOLS.md

File metadata and controls

81 lines (62 loc) · 3.61 KB
Raw

Connect your lab tools

Open Science Desktop is a workbench: it aggregates tools, it doesn't replace them. Anything the agent can reach is an MCP server (Model Context Protocol) or a skill. Both are pluggable — you don't touch app code to add one.

One-click science connectors

Settings → MCP servers lists curated open-source connectors. Click Enable and the app provisions the server into an isolated environment (bundled uv, managed Python — your system is untouched) and registers it. Today:

  • Literature search (all fields) — arXiv, PubMed, Crossref, Semantic Scholar, bioRxiv/medRxiv (paper-search-mcp).
  • Biomedical databases (biology) — PubMed, ClinicalTrials.gov, genomic variants (biomcp).
  • Materials Project (materials) — properties, structures, phase diagrams (mcp-materials-project; free MP API key).
  • FRED economic data (economics) — Federal Reserve time series (fred-mcp; free FRED API key).
  • Space weather (physics) — solar wind, flares, Kp/Dst indices, radiation storms, aurora, from NOAA SWPC / NASA DONKI / USGS (spaceweather-mcp; no key).
  • Weather & climate (earth) — current & historical weather, air quality, timezones from Open-Meteo (mcp-weather-server; no key).
  • USGS water data (earth) — streamflow, flood stages, peak events, sites (usgs-mcp; no key).

Literature and database results carry real identifiers (DOI / PMID / arXiv id), so the traceability-review skill can audit them afterward.

Bring your own MCP server

Any MCP server works — internal ELN, LIMS, a database gateway, an instrument bridge. In Settings → MCP servers, use the add form:

  • local — a command the app launches and talks to over stdio. Example: npx -y @playwright/mcp (browser), or uvx your-lab-mcp for a Python server.
  • remote — a URL the app connects to over HTTP. Example: https://mcp.your-lab.internal/sse.

The entry is written to the bundled OpenCode's config and applies immediately; its live status (connected / failed) shows in the same list.

Minimal local MCP server (Python)

# lab_tools.py — run with: uvx --from fastmcp fastmcp run lab_tools.py
from fastmcp import FastMCP

mcp = FastMCP("lab-tools")

@mcp.tool()
def sample_metadata(sample_id: str) -> dict:
    """Look up a sample in the lab database."""
    return {"id": sample_id, "assay": "RNA-seq", "status": "passed_qc"}

if __name__ == "__main__":
    mcp.run()

Add it as a local server with the command that launches it. Restart-free.

Bring your own skill

A skill is a folder with a SKILL.md (instructions the agent follows) plus any scripts/templates it needs. Install one from the Skills page (paste a URL or Markdown; the agent saves it under the workspace's .opencode/skills/). The app also bundles first-party skills (e.g. traceability-review) and the ai4s-skills pack.

Safety

  • Every server you add can make its own network calls and run its own code — review the source before enabling. The curated list is vetted; your own entries are your responsibility.
  • Command execution, file deletion, dependency installs, and remote connections still go through the agent's approval flow.
  • Provider keys and tokens live in an app-private file, never in the workspace, provenance, logs, or exports.