# Manifestack

> Know what your stack will cost before the bill arrives. Free, MIT-licensed Agent Skill (SKILL.md) for Claude Code, Cursor, Codex and other coding agents. It prices a project's stack at its real usage and growth on vendors' current pages and reports the bills ahead, each with the fix that cuts it. The same audit flags plan limits that cause outages, requirement gaps (EU data, SOC 2), overlapping services and overbuilt setups, each with a when, a cost and a fix. It reads the whole product (frontend, backend, workers, infrastructure code, CI deploy steps, third-party APIs) in JavaScript, Python and Go projects. In a new repo it proposes a stack priced at 1k, 10k and 100k users. Requirements and decisions are kept in `.manifestack/STACK.md`, and every service the agent adds is checked against them. No API keys or account access; it runs inside the agent with no Manifestack server and never changes code, config or vendor settings on its own.

Ask in plain words instead of commands: "what will we pay at 50k users?", "why is our bill growing?", "what breaks first at 50k users?", "what should we build this on?", "Neon or Supabase for this app?".

For developers, small teams and agencies who build with coding agents and choose their own services, from the first commit to a few hundred thousand users.

## Install

Requires Node.js 22 or newer.

- Claude Code: `npx manifestack --agent claude-code`. Then run `/manifestack` in your project.
- Cursor: `npx manifestack --agent cursor`. Then ask the agent to run `manifestack` in your project.
- Codex: `npx manifestack --agent codex`. Then run `$manifestack` in your project.
- Other agents: `npx manifestack`. Detects your agents and asks where to install: Copilot, Windsurf, OpenCode, Cline and more.
- skills CLI: `npx skills add manifestack/manifestack`. Installs the skill without the new-service hook. In Claude Code and Cursor, `/manifestack` offers to add it.
- Claude Code plugin: `/plugin marketplace add manifestack/manifestack`, then `/plugin install manifestack@manifestack`; run it as `/manifestack:manifestack`
- Manual: clone [the repo](https://github.com/manifestack/manifestack) and copy `skills/manifestack` and `skills/manifestack-guard` into `.agents/skills/` (Claude Code: `.claude/skills/`)

The installer adds the skill folders and, for Claude Code and Cursor, one hook entry. `npx manifestack uninstall` removes everything except `.manifestack/`.

The new-service check runs through a hook in Claude Code and Cursor. Other agents have no hook and may not run the check on their own: run `/manifestack audit` before merging a change that adds or swaps a service. The skills CLI installs the skills without the hook; in Claude Code and Cursor, `/manifestack` offers to add it (`npx manifestack hook`).

## Commands

Available now (MVP):

- `/manifestack init`: new project. Asks about budget, users, data rules and what the team prefers and avoids, proposes a stack priced at 1k, 10k and 100k users, writes the first STACK.md
- `/manifestack audit`: existing project. Scans the repo, checks every service against the requirements, budget and growth, lists bills, overbuilt setups, overlaps, limits, requirement gaps and risks, each with a when, a cost and a fix
- `/manifestack compare neon supabase`: two or three vendors for the same job, side by side for the project's workload: requirements, plan and cost at its size, the first limit, the features it uses and what it takes to leave. When one is in use, it also sizes the switch
- Plain `/manifestack` picks `init` for an empty repo, `audit` otherwise, and `compare` when the user asks which of two vendors to use

Coming next:

- `migrate` (step-by-step switch with rollback), `watch` (re-reads vendor pages and proposes a STACK.md diff)

## Data

- Usage numbers come from the user (the skill asks and says where to find each), from exports in the repo, or from a vendor's official MCP connected in read-only mode (for example Supabase, Neon, PostHog). Manifestack never asks for keys or logins.
- It writes only inside `.manifestack/`: `STACK.md`, plus working files in `.manifestack/tmp/` that git ignores. Decisions go into STACK.md (plan, revisit dates) only after the user's yes.
- Runs inside the user's agent session; there is no Manifestack server. It never asks for or sends API keys, env values or billing data.
- The only network calls read vendors' public pages (pricing, limits, regions, certifications) during an audit (needs an agent with web access). Each finding cites the page and the date it was read.

## Vendors with a page map

100+ page maps. AWS, Google Cloud and Azure are each an account map (free tier and credits, spend limits, budgets, data transfer) plus one map per service, so an audit reads only the services a project uses.

- Cloud: AWS (account map plus S3, CloudFront, Lambda, ECS and Fargate (with App Runner), EC2 and Lightsail, Amplify, RDS and Aurora, DynamoDB, ElastiCache, OpenSearch, SES, Cognito, SQS and SNS (with EventBridge, Step Functions), Bedrock, CloudWatch); Google Cloud (account map plus Cloud Run (with functions), App Engine, Compute Engine, GKE, Cloud Storage, Cloud SQL and AlloyDB, Firestore, Memorystore, BigQuery, Pub/Sub (with Tasks, Scheduler, Workflows, Eventarc), Cloud Logging and Monitoring, Firebase); Azure (account map plus App Service and Static Web Apps, Functions, Container Apps, AKS, Virtual Machines, Blob Storage, SQL Database, PostgreSQL and MySQL, Cosmos DB, Redis, Service Bus (with Event Grid, Event Hubs, Queue Storage), Monitor and Application Insights, Entra External ID, Azure OpenAI and Foundry, AI Search, Communication Services)
- Hosting: Vercel, Netlify, Cloudflare, Railway, Render, Fly.io, Heroku, DigitalOcean, Vultr, Hetzner
- Database: Supabase, Neon, Firebase, PlanetScale, MongoDB Atlas, Upstash, Convex, Turso, Prisma Postgres, Pinecone
- Auth: Clerk, Auth0, WorkOS
- Email: Resend, Postmark, SendGrid, Mailgun
- Storage: Cloudinary, UploadThing, Mux
- Payments: Stripe, Paddle, Lemon Squeezy, Polar, PayPal
- SMS and voice: Twilio
- Search: Algolia, Elastic Cloud
- CMS: Contentful, Sanity, Storyblok, Strapi
- Realtime: Pusher, Ably
- Jobs: Inngest, Trigger.dev
- Mobile: RevenueCat, Adapty, Expo EAS, OneSignal
- Monitoring: Sentry, Datadog, New Relic, Grafana Cloud, Elastic Cloud
- Analytics: PostHog, Mixpanel, Amplitude, Segment
- Flags: LaunchDarkly
- AI: OpenAI, Anthropic, Google Gemini, OpenRouter

More ship as releases; request one with the [vendor map issue form](https://github.com/manifestack/manifestack/issues/new?template=vendor-map.yml). For any other vendor, the skill looks up its public pages itself.

No Manifestack backend: everything runs inside your agent.

Fit works in both directions: outgrown plans hit limits and bills, and overbuilt setups (cloud infrastructure built for a million users, serving forty) cost many times what the project needs.

Principle: there's no best stack, only the one that fits a given project (budget, users, data rules, what the team prefers and avoids), and keeps fitting as it grows. The problem is a plan or setup that stopped fitting, not the vendor. Findings are settings to change or upgrades to plan, never rankings or affiliate links.

## FAQ

### Who is it for?

Developers, small teams and agencies who build with a coding agent and pick their own services (Vercel, Supabase, Stripe, AWS and the like), from the first commit to a few hundred thousand users. If nobody on the team owns reading pricing pages, it does that for you. It is not a FinOps tool for large cloud accounts.

### How is this different from asking my agent?

A plain agent answers from training data that can be a year old, then forgets. Manifestack reads today's vendor pages, prices your actual usage, and remembers your requirements in STACK.md, so every later change is checked against them.

### Where do the prices come from?

From each vendor's pricing and limits pages, read during the run, not from training data. It prices your own usage: numbers it finds in the repo, ones you give it, or a vendor's read-only MCP. Every figure cites the page and the date it was read, and a page it can't open is marked unverified instead of guessed.

### Is it only about costs?

No. Bills are where most teams feel it first, so the report leads with them. The same audit flags plan limits that take a site down, data rules such as EU residency or SOC 2, two services doing one job, and setups built for far more users than you have. Each finding has a when, a cost and a fix.

### Will it change my code, config or vendor settings?

No. It is read-only. It writes only inside `.manifestack/`: STACK.md, plus working files git ignores. Decisions go into STACK.md only after your yes.

### Do I need to give it API keys or account access?

No. It reads your repo and env var names, never their values, and asks you for the usage numbers it can't see. If you've connected a vendor's read-only MCP (Supabase, Neon, PostHog), it can read usage there instead.

### Does it work with Cursor, Codex or my agent?

Yes: Claude Code, Cursor, Codex, GitHub Copilot, Windsurf, OpenCode and any agent that reads Agent Skills. `npx manifestack` finds them for you. The automatic check on every new service runs in Claude Code and Cursor; elsewhere, run `/manifestack audit` before merging a new service.

### Does it work with Python, Go, Terraform or mobile apps?

Yes. It reads JavaScript, Python and Go dependencies, config files and env var names, plus infrastructure code (Terraform, CloudFormation and SAM, Bicep, Pulumi) and CI deploy steps. React Native and Expo apps are covered, including RevenueCat, Adapty and EAS. For native iOS and Android it asks about the services it can't see.

### Is it free?

Yes. MIT license, no paid tier, no account. There is no Manifestack server to pay for.

## Links

- [Source on GitHub](https://github.com/manifestack/manifestack): skills, installer and vendor page maps
- [npm package](https://www.npmjs.com/package/manifestack): `npx manifestack` installer
- [Website](https://manifestack.com/): the full page this file summarizes