med-toolkit (Claude Code + Codex)

Medical research tools backed by the med-toolkit remote MCP server. One backend, one set of tools; installs into Claude Code (plugin + skills) or Codex (config.toml + AGENTS.md). The tools are identical on both — only the config wrapper each CLI reads differs.

What you get

Tools (via the med-toolkit remote MCP server — served live, so new and changed tools appear without reinstalling):

• Literature — pubmed_search, semantic_scholar_search, consensus_search, doi_lookup
• Datasets (read-only NHANES / MIMIC-IV / eICU) — list_*_tables, describe_*_table, query_* per dataset family
• Synthetic cohorts — extract_table_to_spec (Table 1 → spec), generate_mock_data (spec → patient-level CSV + fidelity report), plus persistent simulation projects (create/list/get/update_simulation, infer_population, propose_correlations, generate_simulation_data)
• Manuscript sections — write_introduction, write_methods, write_result, write_discussion (outline → cited prose)
• Saved analyses (read-only project graph) — list_analyses, get_analysis, get_dataset, get_analysis_run, get_artifact
• Statistical analysis templates (licensed accounts) — verified, versioned R templates run in a sandbox: list/get/run_analysis_template, extract_dataset, check_template_updates, upgrade_template_pins. Absent from the tool list on unlicensed tokens.

Skills (Claude Code; Codex gets the same guidance via codex/AGENTS.md):

• literature-review — fan out across all three sources, dedupe by DOI, screen, and return a structured, cited reference list
• synthetic-cohort — reconstruct patient-level data from a paper's Table 1 or any per-group summary statistics
• manuscript-sections — draft Introduction / Methods / Results / Discussion from an outline, with verified DOI citations
• statistical-analysis — route figures and tables through the verified R templates (Cox, KM, RCS, Table 1, forest, subgroup, competing risks, chart set) instead of ad-hoc code

Get your API key from the med-toolkit dashboard first (starts with sk_). Without a key, Claude Code can also authenticate interactively — the server answers unauthenticated requests with an OAuth challenge and Claude Code opens the browser login.

Install — full plugin (tools + skills) — recommended

The marketplace is served directly from the med-toolkit backend (no GitHub), and the plugin ships as an npm package. Inside Claude Code:

/plugin marketplace add https://api.med-toolkit.com/plugin/marketplace.json
/plugin install med-toolkit

Set your key (and server URL if non-default) so the connection authenticates:

export MED_TOOLKIT_TOKEN="sk_xxxxxxxx"
export MED_TOOLKIT_MCP_URL="https://api.med-toolkit.com/mcp"   # optional

Restart Claude Code so the MCP server connects.

Update: /plugin marketplace update med-toolkit then /plugin update med-toolkit.

Install — one line, either CLI

The installer detects whichever CLI you have and configures it (both, if both):

curl -fsSL https://api.med-toolkit.com/install.sh | sh -s -- sk_YOUR_API_KEY

Force one with --client claude or --client codex.

• Claude Code — wraps claude mcp add --transport http med-toolkit <url> --header "Authorization: Bearer …".
• Codex — appends an [mcp_servers.med-toolkit] block to ~/.codex/config.toml. Codex speaks MCP over stdio, so the remote HTTP server is bridged with npx mcp-remote. See codex/config.toml for the block, and codex/AGENTS.md for the guidance (Codex's equivalent of the Claude Code skills). Restart Codex.

Update: the tools are a remote server — new or changed tools appear automatically on both clients with no reinstall. Re-run only to change the URL or rotate the key. The skills are baked into the plugin package: they update via /plugin update after a new release.

Verify

Ask Claude: "list the med-toolkit tools" or "find recent RCTs on SGLT2 inhibitors in heart failure" — the literature-review skill should trigger and call the tools.