spectr-mcp 0.1.0

Spectr MCP server — turn a screen recording or App Store URL into a production-ready spec.md, from inside Claude Code / Cursor / any MCP client.

spectr-mcp

MCP server that exposes Spectr's spec-generation pipeline as a tool any MCP client can call. Hand it an App Store URL or a local MP4 of an app, get back a production-ready spec.md precise enough to build a clone from.

See an app. Ship an app.

What it does

One tool, generate_spec(source, ...):

• source — an App Store URL (e.g. https://apps.apple.com/us/app/doordash/id719972451) or an absolute path to a local MP4 screen recording.
• Returns a structured 7-section markdown spec covering app overview, navigation, screens, components, design system (with exact hex/px/weight values), implementation notes, and a Claude Code prompt the developer can paste.
• Targets Expo SDK 54 / React Native / iPhone 15 baseline.

Typical run: 2–5 min for App Store URLs, 5–10 min for MP4 recordings.

Two ways to install

Hosted (recommended) — Claude web app connector

Open Claude → Settings → Connectors → Add a custom connector → paste:

https://mcp.spectr.to

No key, no install, no env vars. The hosted MCP runs on Spectr's Anthropic key; you just paste the URL and start asking Claude to generate specs.

Local (developer install) — stdio via Claude Code

If you have Claude Code installed and logged in (claude login), you can also run the MCP locally as a stdio subprocess. Uses your existing Claude Pro/Max subscription — no API key needed.

Claude Code

claude mcp add spectr -- uvx --from git+https://github.com/Meliwat/spectr spectr-mcp

Or for local development against this repo:

claude mcp add spectr -- uvx --from /path/to/spectr spectr-mcp

Other clients (Cursor, Codex, custom MCP hosts)

Any client that supports MCP stdio servers can launch the local version:

uvx --from git+https://github.com/Meliwat/spectr spectr-mcp

If you don't have the claude CLI installed and authenticated, the MCP needs an ANTHROPIC_API_KEY instead — set it in the launching process's environment.

Or use the hosted version at https://mcp.spectr.to if your client supports remote MCP connectors over HTTP.

Self-hosting the HTTP server

The same package runs the streamable-http transport for hosted deployments:

spectr-mcp --http --port 8000
# Or with env var
SPECTR_MCP_HTTP=1 PORT=8000 spectr-mcp

Set ANTHROPIC_API_KEY in the process env. See ../mcp_server/README.md for the Dockerfile + Railway deployment recipe used to host mcp.spectr.to.

Requirements

• Python ≥ 3.10 (installed automatically by uvx).
• One of:
  • The claude CLI (free path — uses your existing Claude Pro / Max subscription). Install from claude.com/product/claude-code, run claude login once, you're done.
  • ANTHROPIC_API_KEY in env (token-billed path — set it when you want to pay Anthropic directly by token).
• ffmpeg — only required when passing an MP4 path; not needed for App Store URLs.

The MCP auto-selects: API key wins if both are present (faster than CLI subprocess).

Usage from inside Claude Code

> Generate a spec from https://apps.apple.com/us/app/doordash/id719972451 and write it to ./doordash-spec.md

Claude Code will pick the tool, call it, and write the file. The tool returns a short confirmation summary so your context window stays small.

For MP4 input:

> Generate a spec from /Users/me/Desktop/recording.mp4 for the app "Linear"

The reference_app parameter is required for MP4 input (the tool has no way to detect the app name otherwise).

Cost

With Claude Code subscription (CLI path): specs are billed against your Claude Pro / Max plan's normal usage allowance. No per-spec charge from Spectr or Anthropic beyond your existing subscription.

With ANTHROPIC_API_KEY (SDK path): you pay Anthropic by token. Rough numbers per spec at the default max_frames=20:

• Vision (20 frames × 2 passes via Haiku): ~$0.10–0.20
• Spec generation (7 sections via Sonnet, with prompt caching): ~$0.50–1.00
• Total: typically $0.60–1.20 per spec

Raising max_frames past 20 increases vision cost roughly linearly. max_frames=48 (the old default) runs about $1.40–2.00 per spec.

The hosted version at spectr.to (coming back soon) charges $19 per spec and includes the pipeline, hosting, storage, retries, and the realtime progress UI.

License

MIT