The Claude Science toolkit, mapped

A complete inventory of the skills, agent profiles, and MCP data servers behind Anthropic's scientific computing assistant — 29 workflows, 4 agents, and 247 retrieval tools over 62 bio-database packages. Every SKILL.md is readable and downloadable in full.

29Skills

4Agents

24MCP Servers

247Retrieval Tools

62Fleet Packages

science-skills-main claude.com/product/claude-science All skills Apache-2.0

WORKFLOWS

Skills

Domain workflows the agent loads on demand. Each skill is a SKILL.md with YAML front matter — operational procedure, inputs, failure modes, and validation steps. Click View source to read the full file, or Download to save it.

PROFILES

Agent Profiles

Specialized agent personas defined under agents/<name>/metadata.yaml. The orchestrator spawns these for distinct roles. Full YAML is viewable and downloadable.

DATA LAYER

MCP Data Servers

Bundled model-context-protocol servers exposing scientific databases. mcp_bio aggregates all 23 domain servers into one process; the rest are standalone stdio servers backed by accuracy-gated fleet packages.

RETRIEVAL

Fleet Packages

62 Python packages under mcp-servers/bio-tools/lib/ that do the actual API retrieval — pacing, retries, pagination, and honest totals. MCP servers are thin marshalling layers over these.

SYSTEM DESIGN

Architecture

How the layers fit together — from the user's prompt down to the database API call.

1
Agent Core
SYSTEM_PROMPT.md + agents/*/metadata.yaml
The base system prompt defines operating rules; agent profiles (operon, onboarding, reviewer, bookmarker) specialize behavior. The orchestrator decides which skills to load and which MCP servers to spawn.
operononboardingreviewerbookmarker
2
Skills
skills/*/SKILL.md
On-demand workflow packages. YAML front matter declares category, license, GPU requirements, and third-party dependencies (model weights, external services). The body is the operational procedure — commands, gotchas, error tables, validation gates.
biomodels (13)analysis (6)runtime (5)product (4)omics (1)
3
MCP Servers
mcp-servers/bio-tools/ + ketcher-chemistry/
Stdio MCP processes exposing database tools. mcp_bio aggregates all 23 domain servers (247 tools) into one process. Tier-1 servers use captured schemas.json; tier-2 servers use clean FastMCP schemas. All tools are read-only retrieval.
24 servers247 tools+ ketcher-chemistry
4
Fleet Packages
mcp-servers/bio-tools/lib/ (non-mcp_)
62 accuracy-gated retrieval packages that hit the actual public APIs — NCBI, Ensembl, gnomAD, ChEMBL, GTEx, GWAS Catalog, ENCODE, UniProt, and 50+ more. They handle pacing, retries, pagination, and guarantee complete or honestly-truncated results.
62 packages50+ public databases