Configure supported services from your terminal or your AI agent. Stack uses provider-specific OAuth or supplied credentials, provisions where the adapter supports it, and records selected credentials in Phantom Secrets, Stack configuration and provider MCP entries. Vercel requires an existing personal access token; Stack does not create that token. Export an empty .env.example separately with stack env export --example.
Status: pre-alpha, active development. This checkout describes source version 0.2.1; registry metadata checked on October 14, 2026 reported 0.1.1 for the CLI and MCP packages. That published release starts both with Bun, so install Bun first, even for the npx commands below. Check stack --version, stack --help and stack providers for your installed release. Source changes do not establish package publication or provider readiness.
Read-only, no accounts or API keys needed:
npx -y -p @ashlr/stack stack providers # the provider catalog
npx -y -p @ashlr/stack stack recommend "B2B SaaS with auth, AI, and payments" # ranked picks with rationales
npx -y -p @ashlr/stack stack scan # detect services in an existing repo (writes nothing; add --json in agents and CI)To provision a real service, install Stack (below) and run stack add <provider>.
Every MCP tool runs the stack CLI, so the server needs stack on its PATH (or STACK_BIN set to it). These snippets fetch both packages in one npx, which puts stack on the server's PATH:
claude mcp add ashlr-stack -- npx -y -p @ashlr/stack -p ashlr-stack-mcp ashlr-stack-mcp # Claude Code
codex mcp add ashlr-stack -- npx -y -p @ashlr/stack -p ashlr-stack-mcp ashlr-stack-mcp # CodexCursor (~/.cursor/mcp.json):
{ "mcpServers": { "ashlr-stack": { "command": "npx", "args": ["-y", "-p", "@ashlr/stack", "-p", "ashlr-stack-mcp", "ashlr-stack-mcp"] } } }With only ashlr-stack-mcp, the server starts and lists its tools, but every call returns "Could not run the 'stack' CLI".
stack onboard # plan: shows which of Codex, Claude Code, Cursor, Gemini CLI would be wired
stack onboard --apply # write the Phantom gateway entry (backups saved first)
stack onboard --undo # restore the files exactly as they wereOnly the ashlr-hub entry is added; an existing entry that differs is reported and left alone. Use --client gemini,cursor to limit it and --json for scripts.
- One workflow for supported services. Stack authenticates, provisions where supported, and updates vault entries, Stack configuration and MCP wiring.
- Phantom-backed credentials. Stack handles credentials during setup and persists selected entries through Phantom Secrets. Provider requests, project files and CLI output have their own data boundaries.
- Agent-native. MCP exposes selected catalog and setup actions for Claude Code, Codex and Cursor.
Three ways, pick one:
# One-liner, macOS / Linux (also installs Phantom Secrets if missing)
curl -fsSL stack.ashlr.ai/install.sh | bash# One-liner, Windows (PowerShell)
irm https://stack.ashlr.ai/install.ps1 | iex# npm / bun registry
bun add -g @ashlr/stack ashlr-stack-mcp # or: npm i -gDev install (from a local clone of this repo):
git clone https://github.com/ashlrai/ashlr-stack && cd ashlr-stack
bun install
bun run packages/cli/src/index.ts --help
# Optional: alias stack=`bun run $(pwd)/packages/cli/src/index.ts` so it's on your PATHUsing Stack with an AI coding agent? See STACK.md — a self-contained project brief for agents pulling context from the repo.
In a Claude‑Code‑native world, the friction in starting a project isn't writing code — it's tab‑hopping to create and wire ten services. Every npx create-next-app is followed by an hour of:
- creating a Supabase project, copying URL + anon key + service role key into
.env - generating a Vercel token
- picking a Neon region and copying the connection string
- setting up a Sentry project and pasting the DSN
- registering an OAuth app, generating a PAT, pasting keys, adding MCP servers to
.mcp.json…
Stack collapses that hour into one command:
stack init --template nextjs-supabase-posthog
# Run from your application's directory. Stack writes its configuration,
# then provisions template services with provider-specific authentication.
# It does not generate application code or automatically write .env.
stack env export --example # separate empty KEY= skeleton in .env.example- Phantom Secrets — the vault. Stack uses its CLI for selected credential storage and runtime operations; provider authentication still uses credentials.
- Locus — the identity plane. Run
locus run -b <binding> -- stack add …so provider CLIs act only in the pinned personal / company / client account. See docs/phantom-locus.md. - ashlr-plugin — token-efficiency layer for Claude Code. Orthogonal to Stack.
- ashlrcode — multi-provider AI coding CLI. Orthogonal.
Stack is the control plane. Phantom Secrets is the vault. ashlr-plugin is the context compressor. They compose.
Database — Supabase · Neon · Turso · Convex · Upstash · Firebase Deploy — Vercel · Railway · Fly.io · Cloudflare · Render Cloud — AWS AI — OpenAI · Anthropic · xAI · DeepSeek Analytics — PostHog Errors — Sentry Payments — Stripe Code — GitHub Tickets — Linear Email — Resend Auth — Clerk
Run stack providers to see your installed catalog. This source checkout registers 39 providers across 13 categories; the examples above are a subset. stack templates list shows the starter presets. Registration does not establish account access or provider acceptance.
Before stack add, stack apply, or a writing stack import, check
phantom status --json from the project directory. Phantom Secrets 0.7.9
reports initialized: true when its project configuration is readable and valid; this
is metadata, not vault or provider health. For a new project that is not yet
configured, run phantom init --empty in a trusted terminal, then retry.
These write flows stop before provider login or provisioning when initialization
is unknown or missing, and do not initialize Phantom automatically.
stack init # interactive template picker
stack add supabase # provider authentication → setup → vault/config/MCP
stack providers # full provider catalog
stack doctor --fix # inspect the report and attempt supported remediation
stack exec -- bun dev # run with Phantom's secret proxy activeDescribe what you're building — Stack picks the providers.
stack recommend "B2B SaaS with auth, AI, and payments"
# → ranked list of matching providers with rationales
stack recommend "serverless postgres" --save
# → freezes a Recipe to .stack/recipes/<id>.toml
stack apply <recipe-id>
# → provisions recipe providers, then attempts optional Phantom wiring
# and writes webhook stub files; successful rotation is not guaranteed
# (add --noWire to skip the optional wiring layer)Inside Claude Code, the same flow is one tool call:
stack_recommend { query: "B2B SaaS with auth + payments", save: true }
stack_apply { recipe_id: "<id>" }
Reasoning happens in Claude — Stack owns the catalog + execution. Outside Claude, stack recommend --synth uses a local SLM (LM Studio on :1234, Ollama on :11434) for rationales. No remote LLM SDKs live in Stack.
Already have a repo with services wired up? You don't start from scratch.
# In an existing repo:
stack scan # detects Supabase / Sentry / OpenAI / etc. from package.json, config files, .env.example
stack scan --auto # scans, then interactively runs `stack add` for each detection
stack import # or: inhale an existing .env straight into Phantom + .stack.toml
# Clone someone else's project:
stack clone https://github.com/your-org/your-repo
# → git clone + scans the checkout + prints next steps
# Across every project on this machine:
stack projects list # everywhere you've used Stack
stack doctor --all # run health check across all registered projectsStack splits its config into two files so you can commit the shape of a stack without leaking per-developer resource IDs:
.stack.toml— committed. Names of services, their secret slots, MCP wirings..stack.local.toml— gitignored automatically.project_id,resource_id, timestamps. Unique to each clone.
Another developer can inspect stack doctor and use stack doctor --fix for supported remediation. Review its report; existing vault names can cause remediation to be skipped, and metadata presence does not establish valid credentials or a healthy provider.
packages/
core/ — @ashlr/stack-core — shared logic, provider adapters
cli/ — @ashlr/stack — the `stack` binary
mcp/ — ashlr-stack-mcp — MCP wrapper
plugin/ — Claude Code plugin wrapper
site/ — Astro landing page (deploys to stack.ashlr.ai)
templates/ — starter stacks
docs/ — auth matrix, schema reference
Three packages ship to npm: @ashlr/stack-core, ashlr-stack-mcp, @ashlr/stack. For monorepo dev, @ashlr/stack depends on @ashlr/stack-core via workspace:* — that's what lets bun install link the local checkout. A plain npm install @ashlr/stack from outside the workspace cannot resolve workspace:*, so the publish flow has to rewrite those ranges to an actual version (e.g. ^0.1.0) right before npm publish.
Don't hand-edit the workspace:* entries in packages/*/package.json — dev needs them. Release owners should follow docs/RELEASE.md, which describes the current script, its limitations and separate package, native and website acceptance gates.
cd packages/site
bun install
bun run dev # http://localhost:4321
bun run build # static output in dist/Dark-first, magenta accent, Astro + Tailwind v4 + Framer Motion. Three interactive React islands (animated terminal, "with vs without Stack" tab comparison, Claude Code chat mock). prefers-reduced-motion honored.
Ashlr Stack is one of the open-source tools from Ashlr, a forward-deployed engineering company. Its siblings: Lexicon, Locus, Phantom Secrets, Ashlr Markdown, Ashlr plugin.
Issues and PRs are welcome. Look for good first issue to get started, and read CONTRIBUTING.md. Report security problems privately per SECURITY.md.
MIT.