storebound/docs/mcp.md

# MCP server and the desktop machine interface

`storebound mcp` speaks MCP over stdio, tools only: spec revision **2026-07-28** (stateless: `server/discover`,
per-request `_meta`, `resultType`, cacheable `tools/list`) and, on the same process, clients that still open with
`initialize` (2025-11-25, 2025-06-18, 2025-03-26, 2024-11-05). It runs each tool as a CLI call in the project directory
(the server's working directory, or `STOREBOUND_PROJECT`) and returns the output.

- One tool per command. Read-only: `scan`, `check`, `check_aso`, `aso_competitors` (no snapshot saved), `aso_suggest` (never writes), `aso_audit`, `validate`, `report` (writes only the report page in .storebound/reports/ and returns its path and both stores' verdicts), `catalog_diff`, `health`, `reviews_list`, `prices_diff`,
  `analytics_compare`, `apple_status`, `apple_next_build`, `apple_testflight_status`, `apple_ppo_status`,
  `play_status`, `play_next_version_code`, `privacy_manifest`, `shots_check`, `workspace_list`, `workspace_status`,
  `release_status`, `release_plan`, `flow_list`, `flow_run` (dry mode only), `listing_check`, `api_get` (any App Store
  Connect or Google Play GET operation by id). Local writes: `import_fastlane`, `whats_new`, `listing_translate`, `icons`,
  `health_report`, `reviews_draft`, `prices_export`, `analytics_snapshot`, `shots_capture`, `shots_render`,
  `shots_gallery`, `shots_record`, `shots_video`. Store writes: `play_listing`, `play_images`, `play_products`,
  `play_data_safety`, `play_subscriptions`, `play_release_testing`, `apple_bundle`, `apple_iap`, `apple_iap_review`,
  `apple_subscriptions`, `apple_listing`,
  `apple_screenshots`, `apple_previews`, `apple_appinfo`, `apple_review_details`, `apple_encryption`,
  `apple_phased_release`, `apple_testflight_groups`, `apple_testflight_what_to_test`, `apple_testflight_info`,
  `apple_testflight_distribute` (internal groups only), `apple_ppo_create`, `apple_ppo_upload`, `analytics_setup`.
- Annotations say what is read-only, destructive (replaces data), idempotent and network-bound.
- No MCP SDK: a small newline-delimited JSON-RPC server (server/discover, initialize, ping, tools/list, tools/call,
  cancellation), so storebound keeps zero runtime dependencies.

Manual config for a tool that `install` does not know: command `bunx` with args `["storebound", "mcp"]` (or the
absolute path of `bin/storebound` plus `mcp`).

## One shared server over HTTP

`storebound mcp --http` serves the same tools over Streamable HTTP (MCP 2025-06-18 and later) at `/mcp`, so every
agent session of a machine shares one process instead of starting its own. POST carries one JSON-RPC message and gets
one JSON answer; there is no server-initiated stream and no session state. A browser Origin other than localhost is
refused unless `--allow-origin` names it.

- Every tool gets a `project` argument: the absolute path of the app folder to run in. It must be inside `--roots`
  (comma-separated, default the home folder). `--project <dir>` sets a default for calls without one.
- `--auth token` (default): the server makes `~/.storebound/mcp-token` (mode 600) on first start and requires it as a
  bearer token. Only localhost hosts are allowed with it.

  ```bash
  storebound mcp --http --port 7787
  claude mcp add --transport http storebound http://127.0.0.1:7787/mcp --header "Authorization: Bearer $(cat ~/.storebound/mcp-token)"
  ```
- `--auth introspect`: for a server shared over the network behind TLS. Bearer tokens are OAuth 2.0 access tokens from
  your authorization server, checked through its introspection endpoint (RFC 7662; this server authenticates with the
  key in `--introspect-key-file`). A token must be active, carry the `mcp` scope and list `--resource` in `aud`
  (RFC 8707). Without one the answer is 401 with `WWW-Authenticate: Bearer resource_metadata=...`, and
  `/.well-known/oauth-protected-resource/mcp` names `--issuer` (RFC 9728).

  ```bash
  storebound mcp --http --host 0.0.0.0 --auth introspect --introspect-url https://auth.example.com/introspect \
    --introspect-key-file ~/.storebound/introspect.key --resource https://mcp.example.com/mcp --issuer https://auth.example.com
  ```
- `--read-only`: only the tools that change nothing are listed and callable.
- `--broker-profiles a,b`: a host that keeps no key files takes these credential profiles from the credential broker's
  vault (`GET /api/store-credentials/<profile>/asc-api-key | play-service-account`, a key with the `store-credentials`
  scope; docs/credential-broker.md). They reach each tool call as `STOREBOUND_PROFILE_<NAME>_*` env, held in memory for
  10 minutes and never written. A project picks its profile as usual (the config's `account`).
- `project` may also be a folder name under the roots (a shared server keeps its checkouts there).
- Long calls: a `tools/call` whose request accepts `text/event-stream` is answered as SSE, with a comment line every
  20 s until the result arrives as one `message` event, so a proxy that cuts silent responses (Cloudflare: 100 s)
  keeps the connection. Other requests get plain JSON.
- `/health` answers `{ ok, version }` without auth.


## Public key-less server

`storebound mcp --http --public [--port 7788] [--host 127.0.0.1] [--rate 60] [--trust-proxy]` is a separate mode for a
server anyone may connect to (for example behind a public HTTPS address). No sign-in, no token, no project: it never
reads store credentials, project folders or the working directory, never starts a child process and never runs the CLI.
It is read-only and keeps no state. The only network call it makes is the public iTunes Search API (`aso_search`).
It cannot be combined with `--auth`, `--token-file`, `--roots`, `--project`, `--broker-profiles`, `--allow-origin`
or `--read-only`.

Tools (and only these; the CLI tools of the normal server do not exist here, a call gets "Unknown tool"):

- `skills_list`: name and one-line description of every skill in the package.
- `skill_get {name}`: the full SKILL.md of one skill (`^[a-z0-9-]+$`, must be in the list).
- `pitfalls_search {query, store?}`: the sections of the store-pitfalls skill that match the query words, best first (at
  most 8, each clipped). `store` is `apple` or `play`.
- `store_limits {store?}`: listing field character limits, the App Store keyword field rules, screenshot and app
  preview sizes per device class, Play feature graphic and icon sizes.
- `listing_lint {store, fields, locale?}`: storebound's listing and ASO checks on the text you send (limits, keyword
  field rules, repeated name words, other platforms or apps named, placeholder text, keyword stuffing). Apple fields:
  `name`, `subtitle`, `keywords`, `promotionalText`, `description`, `whatsNew`. Play fields: `title`,
  `shortDescription`, `fullDescription`, `whatsNew`. Only the fields you send are judged; nothing is stored.
- `aso_search {term, country?, limit?}`: top apps for a term in one storefront from the iTunes Search API (name,
  developer, id, rating, rating count, genre; 1-25 results, 10 s timeout).

Limits: `--rate N` requests per minute per client (default 60; over the limit the answer is 429 with `Retry-After`),
bodies over 64 KB get 413, every tool input is validated against a closed schema, and output is bounded. Behind a
reverse proxy add `--trust-proxy` so the first `X-Forwarded-For` entry is the client; without it that header is
ignored. Any browser origin may call it (there is no ambient credential), and CORS preflight is answered. `/health`
works as before. Request bodies are not logged.

```bash
storebound mcp --http --public --port 7788 --trust-proxy
claude mcp add --transport http storebound-public https://mcp.example.com/mcp
```

For real store work (builds, uploads, listings, releases) use `npx storebound` and its own MCP server.

## Desktop machine interface

`storebound serve --stdio` is what the desktop app (`apps/desktop`, in progress) runs as its sidecar: newline-delimited
JSON-RPC on the same core as the MCP server (tool list, validation, forbidden flags, child-process runner), with typed
methods (`check`, `validate`, `apple.status`, `play.listing`, ... plus `project.inspect` and `portfolio.status` over
many project folders), structured results (the commands' `--json` output), progress notifications and cancellation.
The first request is `session.hello` with a per-session secret only the host knows. Write methods are dry runs unless
the call carries a single-use confirm token (HMAC over method, arguments and expiry) that the host mints after the
owner confirmed a native dialog showing that dry run. Sending for review is there only as **owner actions**
(`owner.apple_submit`, `owner.apple_resubmit`, `owner.apple_iap_submit`, `owner.play_production`, next to the rollout,
phased release, price and reply owner actions): the window shows the plan (dry run) with the build or versionCode the
owner picked by number, then a native dialog, and only the host's token after that dialog runs it. Nothing publishes,
releases an approved version or cancels a submission, and the MCP server (AI tools) has none of the owner actions.
Play's "Send changes for review" has no API: the Releases tab shows its console plan and a deep link. Types:
`cli/lib/desktop-protocol.ts`.
`bun tools/build-sidecars.ts` compiles the CLI for macOS, Linux and Windows (`bun build --compile`). Design, safety
model and the Dependencies screen (tool sources and licenses): [docs/desktop.md](desktop.md).

---

[All docs](README.md) · [README](../README.md)