DEV Community

linweidao
linweidao

Posted on

Tried `K-Dense-AI/scientific-agent-skills`: Turn Your Coding Agent into an AI Scientist

#ai #cursor #productivity #webdev

Tried K-Dense-AI/scientific-agent-skills: Turn Your Coding Agent into an AI Scientist

K-Dense-AI/scientific-agent-skills is having a serious GitHub moment: +720 stars today. After skimming the repository, the appeal is obvious—it gives general-purpose agents a practical science workflow layer instead of expecting them to improvise research steps.

The project ships 163 validated, ready-to-use skills plus access patterns for 100+ scientific databases across biology, chemistry, medicine, and drug discovery. Think literature review, bioinformatics analysis, molecule lookups, clinical evidence workflows, and reproducible research tasks.

It also follows the open Agent Skills format, so it can plug into tools developers already use: Cursor, Claude Code, Codex, Pi, and Antigravity. That compatibility is likely why it is reportedly used by 175,000+ scientists.

My preferred setup is simple: mount the skills repo into an agent workspace, then route model calls through an OpenAI-compatible gateway.

from openai import OpenAI

client = OpenAI(
    base_url="https://b-lost.com/v1",
    api_key="YOUR_B_LOST_API_KEY",
)

response = client.chat.completions.create(
    model="claude-fable-5",
    messages=[
        {
            "role": "system",
            "content": """
You are a scientific coding agent.
Use the installed scientific-agent-skills repository for validated
biology, chemistry, medicine, and drug-discovery workflows.
Always cite databases, assumptions, and reproducible commands.
""",
        },
        {
            "role": "user",
            "content": "Find candidate drug targets for a KRAS-driven cancer study.",
        },
    ],
)

print(response.choices[0].message.content)
Enter fullscreen mode Exit fullscreen mode

For an agent runner, I’d keep the architecture explicit:

{
  "model": "claude-fable-5",
  "apiBase": "https://b-lost.com/v1",
  "skillPaths": [
    "./scientific-agent-skills"
  ],
  "policy": {
    "requireCitations": true,
    "requireReproducibleSteps": true
  }
}
Enter fullscreen mode Exit fullscreen mode

A relay such as B-Lost is useful here because scientific agents often repeat large system prompts: database rules, skill instructions, citation policies, and experiment context. Its native Anthropic /v1/messages prompt caching can reduce repeated-prompt cost substantially—up to 90% discount on cache hits—while the relay’s listed rate is 0.8× official pricing.

The bigger win isn’t “AI does science automatically.” It’s making research agents follow structured, auditable workflows instead of producing a polished but unverifiable answer.

Top comments (2)

Collapse
 
citedy profile image
Dmitry Sergeev •

idk if turning my coding agent into a full scientist is gonna work or if it'll just get stuck in a loop trying to derive physics constants

Collapse
 
sloves profile image
linweidao •

Accurate! The fine line between "groundbreaking scientific breakthrough" and an agent burning $50 of API credits trying to reinvent Euler's identity at 3 AM is terrifyingly thin. Safe to say human supervision is still very much required!