How many tabs do you have open right now to ship one app?
Play Console · App Store Connect · Firebase · AdMob · Google Cloud IAM...
Write release notes, check screenshot specs, reply to reviews, wire up Firebase — all in separate dashboards, all manual.
Mimi Seed handles all of this through Claude Code or Codex conversation.
"Is my app ready to ship?"
→ Readiness Score 87/100 · 2 blockers: missing 6.9" screenshots, What's New empty
"Write release notes from commits since the last tag, Korean and English, and push to Play Store"
→ 3 tones (concise / detailed / marketing) × 2 locales generated · Apply now?
"Reply to this 1-star review with an empathetic tone"
→ Draft ready · Review before posting?
"Add an Android app to Firebase and download google-services.json"
→ App created · Add SHA-1 fingerprint?
Release Doctor: check Expo, React Native and Unity release risks before connecting app stores.
2026-10-28
ดาว
—
เครื่องมือ
—
อัปไทม์
—
Try Release Doctor first — no account or store key
Run this inside an Expo, React Native, Android, iOS, or Unity repository:
npx mimi-seed check --local
Release Doctor inspects deterministic repository evidence before asking you to connect anything. The first
release checks cover app identifiers, Android Target API requirements, Google Play Billing support dates, the App Store
Connect Xcode / SDK upload minimum (from Xcode versions pinned in CI, fastlane, Codemagic, or eas.json), and legacy
Firebase Cloud Messaging API usage (fcm/send, Instance ID, pre-14.5 firebase-admin topic calls).
The checker is bundled in the CLI, so it does not install the full MCP package or call the network during the
repository-only scan. Every policy finding includes its evidence file, a suggested action, and the official source. Use
--fail-on-blocker in CI, --json for machine-readable output, or --path apps/mobile in a monorepo.
If no Mimi Seed account is connected, a plain npx mimi-seed check automatically takes this local path. After
you connect stores, the same command adds remote listing, screenshot, uploaded-build, and submission-state
checks. A local pass is intentionally not presented as a guarantee of store approval.
Testing an app for a noncommercial purpose? Join the
10-minute Release Doctor pilot. Share redacted finding codes and
timing only—never credentials, private source, absolute paths, or raw JSON.
30-Second Setup
Three steps: 1 Install → 2 Connect → 3 Verify. This page is the short version — for the complete
account setup → CI build → release check → store deploy → social announcement journey, follow the
Mimi Seed User Guide.
Three things that trip every first install:
The local MCP server requires Node 20+.
Open a new session after installing an MCP server or plugin (and after package updates) — tools only appear in fresh sessions.
In Claude Code the 150+ tool schemas load lazily; if a first call fails with InputValidationError, tell Claude: "load the mimi-seed tools with ToolSearch first" (agent guide).
Local MCP — recommended (store-write automation · Google OAuth · runs on your machine, Node 20+)
Claude Code — plugin install (recommended): bundles the MCP server plus the skills that auto-handle deferred tool loading (trap #3) and auth recovery. One fresh session after install is still needed (trap #2):
Remote MCP (status & readiness · requires web console account)
# 1. Create account: https://mimi-seed.pryzm.gg/auth/signin
# 2. Issue a PAT: https://mimi-seed.pryzm.gg/workspace/api-tokens
# 3-a. Register in Claude Code:
claude mcp add --transport http mimi-seed https://mimi-seed.pryzm.gg/api/mcp \
--header "Authorization: Bearer <PAT>"
# 3-b. Or register in Codex:
npx mimi-seed mcp codex --write
# The config references MIMI_SEED_TOKEN instead of storing the PAT.
# Set it in the environment that launches Codex, then restart Codex.
export MIMI_SEED_TOKEN="<PAT>" # bash/zsh
# PowerShell: $env:MIMI_SEED_TOKEN="<PAT>"
What the Remote MCP can and can't do. It exposes a read & diagnostic subset (readiness, blockers, drafts, checklist, publish screenshots) plus workspace-shared BigQuery and App Store IAP review-note/review-screenshot writes. Broader store writes — release-note apply, listing screenshots, Firebase / AdMob / IAM — need the Local MCP (full tool catalog).
2 · Connect your project and accounts
cd <your-app>
npx mimi-seed init # project: auto-detect app → browser sign-in → register apps → agent context files
npx mimi-seed setup # accounts: one guided wizard for stores · CI · social
init detects Expo (including identifiers imported from JSON by app.config.js/.ts) · Gradle · Info.plist · pbxproj · Unity ProjectSettings.asset automatically, and drops .claude/mimi-seed.md plus AGENTS.md so Claude Code and Codex pick up the release workflow every session. (Skipped the plugin and registered the bare MCP server instead? npx mimi-seed init --local chains the Google sign-in and the local-MCP registration into the same pass.)
You don't need every account on day one — most people need two or three. One Google sign-in covers Firebase, AdMob, Play, Google Ads, Search Console, GA4, IAM, and BigQuery; the same wizard also connects App Store Connect, a Play service account, Jenkins, GitHub/GitLab CI, and Facebook / Instagram / Threads whenever you need them → What do you actually need?
On first run setup asks for your language (Enter keeps your system locale: Korean for a Korean locale, otherwise English), then shows what's connected, asks only about what isn't, and skips anything you've already done (so you can quit and resume). At each step press ? to see exactly where to get that token — the full reference is docs/credentials.md. Reconnect the three Meta platforms any time with npx mimi-seed auth meta; switch language with mimi-seed lang en / mimi-seed lang ko (or MIMI_SEED_LANG=en for one command).
App Store Connect: connect once, update safely
npx mimi-seed auth appstore
# Without the CLI:
npx -y @yoonion/mimi-seed-mcp mimi-seed-appstore-auth
The wizard asks for the Issuer ID, Key ID, .p8 path, and an optional Vendor Number. It calls Apple before
saving a new API key; a rejected key leaves the previous working configuration untouched. On an existing
connection, the first prompt has three deliberate paths:
N or Enter: exit without changing anything.
y: replace the primary API key. The saved Vendor Number and a separate reports key are preserved.
v: update only the Vendor Number without re-entering the API key.
Use App Manager for metadata and release work, Admin when creating Analytics report requests, and
Admin / Finance / Sales and Reports for sales reports. Find the digits-only Vendor Number under the Legal
Entity Name at the upper left of App Store Connect Reports. See the
credential reference for the complete permission matrix.
Using the bare MCP server without the CLI? The Google sign-in alone is:
⚠️ Google says "access_denied" / "unverified app"? The OAuth app is in testing mode — only registered test users can sign in, and retrying will never work until you're added. Full recovery → troubleshooting.
3 · Verify, then ask your agent
npx mimi-seed doctor # every credential + the exact fix command for anything missing
Open a new Claude Code / Codex session (trap #2 above) and ask:
"Is my app ready to ship?" — plugin installs can also run /mimi-seed:getting-started for a guided first tour. (The slash prefix follows your server name — a bare registration exposes it as /mimi-seed-local:getting-started; the Remote MCP has no prompts.)
CLI quick reference
Command
What it does
mimi-seed init
Connect project (issue PAT + auto-register apps)
mimi-seed setup
Connect every account, guided — shows what's missing, tells you where each token comes from
git commits → user-friendly release notes → push to stores.
"Write release notes from commits since v2.1.0 in Korean and English, then apply to Play Store"
3 tones: concise / detailed / marketing
Multiple locales in one shot (ko · en-US · ja · zh-TW …)
Generate → review → apply in one flow
AI Review Replies
"Reply to this 2-star review with a professional tone"
Tones: friendly · professional · empathetic · brief
Post with playstore_reply_review after review — the first call (without confirm) returns a dry-run preview; the reply is published only when you call again with confirm: true.
AI-generated replies are drafts. Always review before posting.
Screenshot Spec Validation
Validate local files against store requirements before uploading.
"Do these screenshots meet the iPhone 6.9-inch spec?"
Works with Jenkins · GitHub Actions · GitLab CI (auto-detected, or force with --ci).
Team-Shared BigQuery (Remote MCP)
Give a whole team read-only BigQuery access (GA4 export analysis, etc.) with one shared
service account — no per-machine key files, and immune to the Google Workspace OAuth
reauth policy (invalid_rapt) that breaks personal tokens.
A workspace owner/admin registers the service account once (encrypted, workspace-scoped):
"Register a BigQuery service account for the workspace"
→ register_integration(provider="bigquery", key="serviceAccountJson", value=<SA key JSON>)
→ register_integration(provider="bigquery", key="projectId", value="my-gcp-project")
Then every workspace member (invite them at /workspace/members, they issue a PAT at
/workspace/api-tokens) connects the Remote MCP and queries with no local key:
"List BigQuery datasets" # bigquery_list_datasets
"Run: SELECT COUNT(*) FROM `proj.ga4.events_*` WHERE ..." # bigquery_run_query (SELECT/WITH only)
Members can use the SA for read-only queries but cannot read the key back (no tool
returns stored secret values), and write statements are blocked. Recommended IAM on the SA:
roles/bigquery.jobUser + roles/bigquery.dataViewer.
The Local MCP also has bigquery_* tools, but those authenticate with your own~/.mimi-seed key/OAuth — the shared SA lives on the Remote endpoint.
Project Manifest — per-teammate setup guidance
Drop a .mimi-seed.json at your repo root declaring the services the project needs.
mimi_seed_status (MCP) and mimi-seed doctor (CLI) read it and tell each teammate exactly
what they're missing + the precise setup command — instead of a generic scan.
socialProfiles selects named local credentials from
~/.mimi-seed/social-profiles/<profile>.json. A profile file may hold both instagram and threads
credentials, and different projects can map either platform to a different profile. When no mapping is present,
the legacy ~/.mimi-seed/instagram.json and threads.json files remain the defaults. Create or refresh a named
profile with mimi-seed auth instagram --profile my-app and mimi-seed auth threads --profile my-app.
A teammate who clones the repo just runs mimi-seed doctor (or asks the agent "what am I
missing?") and follows the ❌ items. bigquery reports honestly whether a service account
or OAuth fallback is present.
Slash Commands (MCP Prompts)
Available in any MCP client (Claude Code, Codex, etc.) as native slash commands:
Command
What it does
/mimi-seed:getting-started
First-run onboarding — connection scan → capability tour → first read-only action
/mimi-seed:deploy
Check blockers → generate release notes → apply to stores
/mimi-seed:health
Auth status + launch readiness summary
/mimi-seed:review-inbox
Fetch unanswered reviews → draft AI replies
Plus MCP resources: mimi-seed://auth/status (token state) · mimi-seed://agent/guide (the full agent guide, served over MCP) · mimi-seed://tools/catalog (every tool by domain, with the credential each domain needs).
Skills & Agent Guide
Bundled skills (auto-loaded by the Claude Code / Codex plugin from skills/):
Skill
Use it for
mimi-seed
General entry — status → readiness → release notes → store apply
mimi-seed-onboarding
First run — install check → what can it do → minimal credentials → first safe action
playstore-publish
Play Store listing, images, release notes, track promote
appstore-publish
App Store Connect What's New + screenshots
deploy
End-to-end: CI build → blocker check → notes → apply
video-create-publish
Story/Shorts production → typography and crop review → render validation → optional YouTube publish
mimi-seed-install
Install & register from a git checkout — contributors / unpublished code
mimi-seed-update
Update to the latest server / skills / CLI — and verify the running version
Building an agent on top of Mimi Seed? Read docs/agent-guide.md —
how tools load (the deferred-tool / ToolSearch select: pattern), the right call order,
auth model, and which actions are irreversible.
Contributing to the SDK? The domain ontology in docs/domain/ maps the architecture, the
full tool catalog, the auth/credential model, and known pitfalls — start at
docs/domain/_index.md.
Local MCP Tool List (150+ tools · 23 domains)
These run via the Local MCP — Google OAuth on your machine. The Remote MCP exposes a smaller read/diagnostic subset plus App Store IAP review-note/review-screenshot writes. Always-current catalog: docs/domain/tool-catalog.md. Every tool carries MCP annotations (read-only / destructive / open-world), and destructive tools (submit, delete, public posts, …) return a dry-run preview unless called with confirm: true.
Server base URL (default: https://mimi-seed.pryzm.gg)
ANTHROPIC_API_KEY
Enable AI release notes and review replies (optional)
YOUTUBE_API_KEY
Reference-video metadata research for video_research_youtube (optional)
PEXELS_API_KEY
Licensed stock-video search for video_search_stock_assets (optional)
OPENAI_API_KEY
Generated scene images for video_generate_image (optional)
MIMI_SEED_FFMPEG_PATHMIMI_SEED_FFPROBE_PATH
Optional absolute executable paths when FFmpeg/ffprobe are not on PATH
MIMI_SEED_LANG
Force CLI output language (ko / en) — wins over ~/.mimi-seed/settings.json; with neither set, a Korean system locale gets Korean and any other locale English
MIMI_SEED_TOOLSETSMIMI_SEED_TOOLSETS_EXCLUDE
Local MCP only: expose just these tool domains (comma-separated domain keys such as playstore,appstore, or groups store · google · social · media · build · all). Unset = every domain. auth and checks are always on; unknown keys are ignored with a warning. Keeps tool lists small in clients without lazy tool loading (agent guide)
Bring your own Google OAuth client. Otherwise it is fetched from the web console at login — set these if you're offline, air-gapped, or self-hosting (troubleshooting)
Legacy Compatibility
Data from the Preseed era (~/.preseed/) is picked up automatically — no re-auth needed.
Reads ~/.preseed/tokens.json and ~/.preseed/appstore.json if present
Still honors PRESEED_GOOGLE_CLIENT_ID / PRESEED_GOOGLE_CLIENT_SECRET