Skip to content

STRATEGY.md

Latest commit

 

History

History
137 lines (93 loc) · 9.33 KB

STRATEGY.md

File metadata and controls

137 lines (93 loc) · 9.33 KB
Raw

Stack — Strategy

Last updated: 2026-04-19. Supersedes scattered notes; authoritative until the next rev lands here.

What Stack is (in one line)

The control plane that turns one Claude Code prompt — "stack my backend" — into a running stack with real accounts, vaulted secrets, and MCP wiring.

The problem, framed for a term sheet

In a Claude-Code-native world, writing code is no longer the bottleneck. The bottleneck is account sprawl: every new project requires 10+ third-party services (database, auth, payments, analytics, errors, deploy, AI), each with its own dashboard, token flow, and .env incantation. A senior engineer burns 60–90 minutes per project on account creation, secret copy-paste, and wiring — before writing a single line of product code.

Existing tools solve slices:

  • Vercel v0 — UI code generation only (no services).
  • Railway — deploy only (no upstream provider accounts).
  • Supabase AI — Postgres schema only (single-provider lock-in).
  • Pulumi AI — IaC code generation (doesn't create accounts, doesn't pick providers).
  • Doppler / Phantom — secrets only (you've already picked your providers).

Nobody owns "zero to running prod stack with informed provider selection." That's the wedge.

Distribution wedge

Three channels, all live on day one:

  1. npm i -g @ashlr/stack — dev-tool default.
  2. brew install ashlrai/ashlr/stack — homebrew-ashlr tap, mirrors Phantom's distribution.
  3. Claude Code MCP — stack_recommend + stack_apply tools, native in every Claude Code session. One slash-command install via the ashlr-stack plugin.

The MCP channel is the asymmetric bet: every Claude Code user who says "build me a SaaS" is a latent Stack customer. The CLI is the surface; MCP is the agentic-workflow conversion.

Product shape today (v0.1)

  • CLI (@ashlr/stack): 22 commands. init, add, recommend, apply, scan, clone, status, doctor, exec, sync, open, info, providers, templates, env, deps, remove, import, login, projects, completion, upgrade, ci.
  • MCP server (ashlr-stack-mcp): 19 tools. Same surface, agent-addressable.
  • Claude Code plugin (ashlr-stack): /stack:recommend, /stack:apply, /stack:add, /stack:doctor, /stack:list, /stack:status.
  • Catalog: 30 working CLI adapters (27 from the AI-layer session + 3 email adapters landed tonight: SendGrid, Mailgun, Postmark). Site currently advertises 39; the remaining 9 phantom entries carry "coming soon" semantics.
  • AI layer: stack recommend "<query>" → BM25+IDF retrieval over the catalog. --save freezes a Recipe. stack apply <id> replays through stack add for every provider + pre-wires Phantom rotation envelopes + drops webhook stubs (Stripe, Clerk, Supabase, GitHub). Claude-MCP-first reasoning; LM Studio/Ollama local SLM fallback for standalone CLI.
  • Phantom integration: every secret value lives in Phantom. Stack only holds slot names in .stack.toml. Zero vault logic in Stack.
  • Thick lifecycle: stack add supabase creates a Supabase project via the Management API. Same pattern for Neon, Vercel, Railway, Sentry, PostHog, GitHub, etc.

The moat (three layers, ordered by ship-time)

Moat 1 — Phantom auto-rotation + webhook pre-wiring ✅ shipped

When the AI picks Stripe + Supabase + Clerk for a SaaS, Stack asks Phantom to create rotating secret envelopes for every declared key + drops webhook-endpoint stubs. This requires owning a secrets product; Railway/Supabase/Pulumi can't copy it without building Phantom.

Moat 2 — Live provider MCP queries (next)

stack recommend calls Stripe/Supabase/Vercel MCPs at recommend-time to check current pricing, region availability, quotas. Synthesis becomes: "Stripe Billing ($29/mo) + Supabase EU ($10/mo) + Vercel Edge (dynamic) — total ≈ $40/mo for your scale." Competitor copy = maintaining 15+ partner integrations.

Moat 3 — Cohort telemetry (post-scale)

Anonymized "stacks chosen together" signal feeds the recommender. Requires users first; becomes a network effect ~10k projects in.

Business model

Free OSS CLI + paid cloud tier for teams. Target $20–$200 / mo / team.

Free (MIT, forever)

  • All 22 CLI commands, all MCP tools, all provider adapters.
  • Local-first. No account required.
  • Phantom-backed secret rotation (uses the user's own Phantom install).
  • Claude Code plugin.

Paid (stack cloud) — v0.2+

Tier Who What Price
Team 2–20 developers Shared recipe marketplace (org-scoped); team-sync so one stack cloud sync provisions everyone's local .stack.toml + Phantom envelopes; audit log of who provisioned what; Stack-managed OAuth apps (skips the "register your own Supabase OAuth client" problem); usage analytics per provider. $29/user/mo
Business 20–500 developers Team + SOC2 audit logs, SSO (via WorkOS), private catalog extensions (BYO provider adapters), priority support, Slack/Discord shared channel. $99/user/mo
Enterprise 500+ developers Business + on-prem recipe marketplace, dedicated infrastructure, SLAs, security review, custom contracts. Contact sales

Revenue-triggering metrics

  • Time-to-first-deploy — stack-recommended-then-applied → first service live. Target <10 min.
  • Stacks provisioned/day — proxy for adoption velocity. Target: 1k stacks/day by M6.
  • Provider partnership depth — closed OAuth integrations (Supabase, Vercel, Stripe, GitHub). Each = ~5-second account-creation on behalf of the user.
  • Cross-product adoption — % of Stack users who also install Phantom Cloud, ashlrcode, ashlr-workbench. Measures platform-thinking.

ARR path

  • M1–M3: 500 free users (GitHub stars + HN launch).
  • M6: 5k free users, 50 paid teams ($29k MRR, $348k ARR).
  • M12: 25k free users, 500 paid teams ($290k MRR, $3.5M ARR).
  • M18: 100k free users, 2k paid teams + 5 enterprise ($1.2M MRR, $14M ARR).

These are pre-Series-A targets. A16z / YC infrastructure bar for Series A ≈ $2M ARR with 15%+ MoM growth.

Roadmap (honest, dated)

v0.1 (shipped 2026-04-19)

  • CLI + MCP + plugin installable via npm / brew / curl.
  • 30 CLI adapters + 9 site-advertised "coming soon" entries.
  • AI recommendation layer (stack recommend) + stack apply with Phantom auto-rotation.
  • Golden path: auto-init + partial-failure rollback.
  • Full AI discoverability (openapi.json, mcp.json, llms.txt, /docs/recommend).
  • 186 tests, tsc clean, biome clean.

v0.2 (target 2026-05)

  • Ashlr-managed OAuth apps so users don't register their own Supabase / GitHub OAuth clients (kills the biggest first-install friction).
  • Headless OAuth fallback (SSH / CI detection → PAT prompt instead of browser timeout).
  • stack cloud login + basic account plumbing.
  • Close the 9 remaining phantom providers to catalog parity (Auth0, WorkOS, Mixpanel, Plausible, Datadog, Grafana, GCP, DigitalOcean, Hetzner, LaunchDarkly).
  • Recipe marketplace (read-only, community-curated, pre-vetted).
  • 60-second install-to-deploy demo video for the landing page.

v0.3 (target 2026-06)

  • Live provider MCP queries (Moat 2): real-time pricing + region + quota synthesis.
  • Paid Team tier goes live.
  • Cross-provider secret-rotation orchestration (one command rotates every key in a stack).
  • VS Code + JetBrains extensions (wrappers around the CLI + MCP).

v0.4+ (target 2026-Q3)

  • Cohort telemetry (Moat 3).
  • Business tier (SOC2, SSO, audit log).
  • Stripe Connect OAuth for the marketplace.

What NOT to do

  • Do not build vault logic in Stack. Phantom owns secrets; Stack owns slot names. Always compose.
  • Do not add remote-LLM SDKs to the Stack repo. (openai, @anthropic-ai/sdk, etc.) The user's Anthropic key belongs to Claude, not to us. OpenAI-compatible local endpoints only.
  • Do not introduce a community recipe DSL in v0.x. Curated top-30 catalog hand-tuned beats a Flatpak-style marketplace pre-scale.
  • Do not ship provider adapters without a verify() call. Every adapter either hits a real Management API endpoint on stack add or explicitly documents why it's shape-check only (Modal, Convex).
  • Do not ship anything on the site that the CLI can't serve. The "phantom providers" gap was tolerable for two days; it isn't tolerable for v0.2.

Open strategic questions

  • Ashlr OAuth app registration. Currently blocking because OAUTH_DEFAULTS.supabase = "". Unblocks browser-OAuth for Supabase + GitHub and eliminates 90% of first-run friction. Highest-leverage outside-code work.
  • Homebrew tap split vs. unify. Phantom tap exists (homebrew-phantom). Should Stack ship under the same tap (homebrew-ashlr) with Phantom re-tapped, or stay separate? Recommend: unify — single brew tap ashlrai/ashlr covers the whole suite.
  • Telemetry opt-in defaults. Cohort-telemetry moat needs data. Default opt-in (sharper moat, weaker trust) vs. default opt-out (lower signal, better launch narrative)? Recommend: opt-out with a compelling "share anonymous provider picks to improve the recommender" in-CLI prompt on first stack recommend.
  • VC timing. Current state (v0.1 launching, MIT, no revenue, sibling products in Ashlr family) is raisable on narrative alone. Delay until v0.3 for real ARR numbers? Recommend: talk to 5 funds at v0.1 for term-sheet optionality; close at v0.2 when the demo video is public and the first 10 paid teams are live.