Supstack Cli MCP Server

@supstack/cli

Evidence-based supplement intelligence in your terminal — and an MCP server for AI agents.

A thin client over the public SupStack API. Read-only, no account required. One capability registry powers both the CLI and the MCP server.

Status: Phase 2 (accounts) in progress. All read commands below work, plus an MCP server and login/whoami. Install with npm install -g @supstack/cli.

Install

# Homebrew (macOS / Linux)
brew install drbaher/supstack/supstack

# or npm
npm install -g @supstack/cli

# or run without installing
npx -y @supstack/cli define adaptogen

Then enable shell completion: supstack completion install.

From source (in this directory): npm install && npm run build && node dist/index.js define adaptogen.

Usage

supstack research magnesium --protocol       # full evidence summary for one supplement
supstack search --goal deep-sleep -n 5       # search by name or filter
supstack compare magnesium glycine           # 2–3 head-to-head
supstack studies "sleep" --type rct          # research library
supstack interactions caffeine l-theanine --pathway   # interaction check (deep pair analysis)
supstack interactions magnesium potassium --medication lisinopril   # supplement × drug check
supstack stack add magnesium --dose 400mg --timing bedtime   # local stack (+ dose/timing/brand)
supstack stack sync                          # sync local ⇄ your account (login required)
supstack goals --category sleep              # list goal ids (for rate/recommend)
supstack rate --goals deep-sleep,sharpen-focus   # grade your stack A–F by goal coverage
supstack experiments start magnesium deep-sleep  # run an N-of-1 experiment (requires login)
supstack export --format md                  # export your stack
supstack define bioavailability              # glossary lookup
supstack <command> --json                    # machine-readable output on any command
supstack --help

Account (Phase 2)

supstack login          # sign in via device-code flow (opens the browser to confirm)
supstack whoami         # show the signed-in account
supstack logout         # sign out and revoke this device's token

login shows a one-time code, opens https://supstack.me/activate, and finishes once you approve in the browser. The token lives in ~/.supstack/config.json (0600). Reads work fully anonymously — an account unlocks personalized features.

Once logged in, sync your stack with your account:

supstack stack pull     # local ← your account
supstack stack push     # local → your account (keeps dosage/timing/brand for kept items)
supstack stack sync     # additive merge of both; preserves existing cloud metadata

supstack profile                              # view your health profile
supstack profile set --age 35 --sex male --weight 80 --weight-unit kg
supstack recommend                            # personalized picks from your goals + stack
supstack experiments list                     # your N-of-1 experiments + verdicts
supstack track log                            # log today's stack as taken
supstack track adherence                      # your adherence rate, streak, per-supplement

As an MCP server

supstack mcp   # stdio MCP server exposing all capabilities as tools

This gives an agent the full SupStack toolset. Read-only (no account needed): supstack_research, supstack_search, supstack_compare, supstack_studies, supstack_interactions, supstack_stack, supstack_rate_stack, supstack_export, supstack_define.

Account-scoped tools are also exposed — supstack_recommend, supstack_profile_get / supstack_profile_set, supstack_experiments_list / supstack_experiments_get, supstack_track_log, supstack_track_adherence (plus cloud supstack_stack pull/push/sync). These require the user to be signed in (supstack login, or a SUPSTACK_TOKEN); without a token they return a clear "not logged in" error rather than failing. Mutating tools (profile_set, track_log) are flagged readOnlyHint: false for the agent.

Claude Code (one command):

claude mcp add supstack -- supstack mcp

Claude Desktop — add to claude_desktop_config.json (macOS: ~/Library/Application Support/Claude/claude_desktop_config.json):

{
  "mcpServers": {
    "supstack": {
      "command": "supstack",
      "args": ["mcp"]
    }
  }
}

Or run it without a global install via "command": "npx", "args": ["-y", "@supstack/cli", "mcp"]. To use a local checkout, point at the built entrypoint — "command": "node", "args": ["/absolute/path/to/supstack-cli/dist/index.js", "mcp"] (run npm run build first).

Configuration

Global flags available on any command: --json, --no-cache, --timeout <seconds>, --color / --no-color, -q, --quiet.

supstack auth set-key <key> persists a key to ~/.supstack/config.json (written 0600).

Response cache

Read-only API responses are cached under ~/.supstack/cache/ (1-hour TTL by default) to keep repeat lookups well under the 60/min rate limit. Bypass it per command with --no-cache, or manage it with supstack cache clear / supstack cache path. The cache is bounded (oldest entries pruned past a cap).

Shell completion

One-step install for your current shell (writes the script and wires your rc file idempotently; uninstall reverses it):

supstack completion install      # detects bash | zsh | fish from $SHELL

…or generate the script yourself:

supstack completion bash >> ~/.bashrc
supstack completion zsh  > "${fpath[1]}/_supstack"
supstack completion fish > ~/.config/fish/completions/supstack.fish

Completions are dynamic — the script forwards what you've typed to supstack __complete, which offers the right thing for the position:

• top-level commands and sub-actions (stack <TAB> → add remove list pull push sync)
• supplement slugs where a slug is expected (research <TAB>, compare a <TAB>, stack add <TAB>, track log <TAB>)
• goal ids after search --goal <TAB>

Slug/goal lists are fetched once from the API and cached under ~/.supstack/completion/ (24-hour TTL). Pre-warm or refresh them with:

supstack completion refresh

Exit codes

Commands exit with a semantic code so scripts and MCP wrappers can branch on the kind of failure:

Develop

npm test            # unit tests (mocked fetch)
npm run type-check  # tsc --noEmit
npm run build       # tsup → dist/

See CLAUDE.md for the capability pattern and how to add a command.

License

MIT