AI recommend + apply

Why this exists

Picking the right providers is the first friction point in a new project. Stack's catalog has 39 curated providers across 12 categories — too many to hold in your head, too few to outsource to a generic LLM. stack recommend closes the gap: free-text in, ranked providers out, frozen to a reproducible Recipe that stack apply can replay.

Quick start

Three lines from a blank repo to a provisioned stack:

stack recommend "B2B SaaS with auth, AI, and payments" --save
# → .stack/recipes/b2b-saas-with-auth-ai-and-paymen.toml

stack apply b2b-saas-with-auth-ai-and-paymen

Two modes

Stack runs in whichever mode matches where you're typing. Neither mode ever calls a remote LLM from Stack itself.

stack recommend

Rank providers against a free-text description. Flags compose.

Default output

$ stack recommend "B2B SaaS with auth, AI, and payments"

┌   stack recommend
  query: B2B SaaS with auth, AI, and payments

  ● Clerk  Auth  (4.13)
    Drop-in auth + users. Secret key stored in Phantom.
    add with: stack add clerk

  ● Stripe  Payments  (3.70)
    Billing, subscriptions, tax. Restricted / test-mode secret key stored in Phantom.
    add with: stack add stripe

  ● xAI  AI  (2.90)
    Grok + tool use. Key verified on paste.
    add with: stack add xai

  ● Sentry  Errors  (2.48)
    Error + performance tracking. Creates a project and fetches its DSN.
    add with: stack add sentry

  ● OpenAI  AI  (2.23)
    GPT, Realtime, embeddings. Key verified against /v1/models on paste.
    add with: stack add openai

  ● Anthropic  AI  (2.23)
    Claude models + MCP. Key verified against the Messages API on paste.
    add with: stack add anthropic

  Top pick per category: Auth→clerk, Payments→stripe, AI→xai, Errors→sentry.
  No matches surfaced for: Database, Deploy, Cloud, Analytics, Code, Tickets, Email.
  Apply a chosen set with: stack add clerk stripe xai (or supply your own list).

  Tip: rerun with `--save` to freeze as a recipe for `stack apply`.
│
└  Tell Claude or run `stack add <name>` to continue.

--save — freeze to a Recipe

$ stack recommend "B2B SaaS with auth" --save

  ● Clerk  Auth  (4.13)
  ● Sentry  Errors  (2.48)
  ● Supabase  Database  (2.16)
  ● Firebase  Database  (2.16)
  ● Turso  Database  (0.79)

  Saved recipe: b2b-saas-with-auth   (/path/to/.stack/recipes/b2b-saas-with-auth.toml)
  Apply with:  stack apply b2b-saas-with-auth

--synth — local-SLM rationales

With LM Studio or Ollama running, each hit gains a why: line authored by the model. Without either, the command behaves exactly like the default — no prompts, no failures:

$ stack recommend "analytics + error tracking for a Next.js app" --synth

  ● PostHog  Analytics  (5.01)
    Product analytics + session replay. Personal API key stored in Phantom.
    why: PostHog gives you funnels, flags, and replay for a Next.js app in one SDK.
    add with: stack add posthog

  ● Sentry  Errors  (4.77)
    Error + performance tracking. Creates a project and fetches its DSN.
    why: Sentry captures unhandled exceptions and Web Vitals without extra glue.
    add with: stack add sentry

--json — machine-readable

Same shape the MCP server emits back to Claude. Good for scripts, CI, and your own tooling:

$ stack recommend "B2B SaaS with auth and payments" --json
{
  "query": "B2B SaaS with auth and payments",
  "hits": [
    {
      "name": "clerk",
      "displayName": "Clerk",
      "category": "Auth",
      "authKind": "api_key",
      "secrets": ["CLERK_SECRET_KEY"],
      "blurb": "Drop-in auth + users. Secret key stored in Phantom.",
      "score": 4.125,
      "matched": ["auth"]
    },
    {
      "name": "stripe",
      "displayName": "Stripe",
      "category": "Payments",
      "authKind": "api_key",
      "secrets": ["STRIPE_SECRET_KEY"],
      "blurb": "Billing, subscriptions, tax. …",
      "score": 3.697,
      "matched": ["payments"]
    }
  ],
  "byCategory": { "Auth": [{ "name": "clerk", "score": 4.125 }], "…": [] },
  "guidance": "Top pick per category: Auth→clerk, Payments→stripe. …"
}

--k — cap the hits

$ stack recommend "postgres database" --k 3

  ● Neon  Database  (5.76)
    Serverless Postgres. Creates a project and pools the connection string.
  ● Supabase  Database  (5.49)
    Postgres + Auth + Storage. Full upstream provisioning via the Management API.
  ● Turso  Database  (4.35)
    Edge SQLite (libSQL). Creates a database in your default org.

--category — restrict to one category

$ stack recommend "serverless" --category Database

  ● Neon  Database  (5.76)
    Serverless Postgres. Creates a project and pools the connection string.
  ● Turso  Database  (2.18)
    Edge SQLite (libSQL). Creates a database in your default org.

stack apply

Replay a saved Recipe through the existing stack add pipeline, then pre-wire Phantom rotation so the first provisioned credential lands into an already-rotating envelope.

$ stack apply b2b-saas-with-auth

  recipe   b2b-saas-with-auth
  query    B2B SaaS with auth
  providers clerk, sentry, supabase

  › stack add clerk
  › stack add sentry
  › stack add supabase

  envelopes 3 · webhooks 2

  › stack doctor
  ✓ b2b-saas-with-auth applied.

What gets wired

With the default flow (no --noWire) apply also writes:

• Phantom rotation envelopes — one per secret slot the recipe's providers declare. The envelope rotates on a cadence from day one, so the credential you just fetched is never the credential that ships to prod a week later.
• Webhook stubs — scaffolded verification + handler files for any provider the recipe includes with a known rotation/webhook contract: stripe, clerk, supabase, github. Drop them into your app; they're ready to verify signatures against Phantom-held secrets.
• Inline stack doctor — runs at the end so you see any red dots before apply exits.

MCP tools

Both commands are exposed as MCP tools so Claude Code (and any MCP-speaking editor) can drive the full loop from a coding session. See the full MCP integration docs for setup.

In-session example

A real back-and-forth once Stack's MCP server is installed in Claude Code:

You > Build me a SaaS — auth, payments, error tracking, Postgres.

Claude > I'll pick the providers first and freeze a recipe so you can
         replay this.
         (calls stack_recommend { query: "SaaS with auth, payments,
          error tracking, and Postgres", save: true })

         Top picks: clerk (Auth), stripe (Payments), sentry (Errors),
         neon (Database). Saved as recipe "saas-with-auth-payments-error".

         Want me to apply it now? This runs `stack add` for each and
         pre-wires Phantom rotation + webhook stubs.

You > Yes.

Claude > (calls stack_apply { recipe_id: "saas-with-auth-payments-error" })

         ✓ 4 providers provisioned · 7 rotation envelopes · 3 webhook stubs
         All services healthy (stack doctor).

Recipe file format

Recipes are plain TOML under .stack/recipes/<id>.toml. Safe to commit — provider names are stable, the file contains no secrets (credentials live in Phantom).

id = "b2b-saas-with-auth"
query = "B2B SaaS with auth"
createdAt = "2026-04-19T02:22:44.810Z"
guidance = "Top pick per category: Auth→clerk, Errors→sentry, Database→supabase."

[[providers]]
name = "clerk"
rationale = "auth"

[[providers]]
name = "sentry"
rationale = "auth"

[[providers]]
name = "supabase"
rationale = "auth"

The id is the first ~40 chars of the query, slugified. The rationale starts as the matched terms from retrieval and gets overwritten when you re-run with --synth or when Claude synthesizes via MCP.

How retrieval works

Zero-dep BM25 + IDF scorer over the curated provider catalog, with a small table of hand-authored synonyms (db → database, billing → payments, auth → users, …) so natural-language queries map to the terms we actually index. Scoring is deterministic — same query, same ranking. The catalog itself lives in packages/core/src/catalog.ts and is the same source the providers page renders from.

FAQ

Can I use my own LLM?

Yes. Set STACK_LM_URL to any OpenAI-compatible endpoint. Stack talks plain HTTP to LM Studio (localhost:1234) and Ollama (localhost:11434) out of the box — no SDKs, no API keys in the Stack process.

Does anything call Anthropic or OpenAI?

No. All synthesis is either local (your SLM) or delegated to Claude Code via MCP. Stack's repo has zero remote-LLM SDKs. The user's LLM keys belong to Claude, not to us.

What if the SLM isn't running?

--synth silently falls back to retrieval-only output. The command still exits 0, the ranking is unchanged, and --save writes the same Recipe shape. No prompts, no errors — retrieval alone is the floor, synthesis is the bonus.

What now?