J Quants MCP

jquants-mcp

English | 日本語

An MCP (Model Context Protocol) server that retrieves Japanese stock market data via J-Quants API v2.

User-facing documentation site: https://shigechika.github.io/jquants-mcp/ (also available in 日本語) — start there if you want a gentler 5-minute introduction. This README is the technical reference (config schema, all 55 tools with parameter tables, deployment).

Release history and changelog: GitHub Releases.

Deployment shapes and how to pick between them: see docs/deploy/.

Demo

24-second loop showing real output from the Claude iPhone app calling jquants-mcp tools:

• Sector performance ranking (業種別騰落率) — get_sector_performance
• Top turnover by trading value (売買代金ランキング) — get_top_turnover_value
• Candlestick chart with SMA — get_candlestick_data
• Quarterly financial summary (決算ダイジェスト) — get_fins_summary
• 5-stock return comparison — get_comparison_chart_data

Individual frames are in docs/screenshots/.

Features

• 55 MCP tools — 22 J-Quants API v2 endpoints, 11 market overview + valuation, 10 offline screener, 1 technical indicators, 1 single-stock summary, 3 cache-only equity search + earnings (schedule + results), 2 chart tools (JSON, no optional dependencies), and 5 server utilities
• Two-tier SQLite cache — row-level cache for time-series data, response-level cache with TTL for others
• Stock split detection — automatic cache invalidation when AdjFactor changes
• Rate limiting — plan-aware sliding window (Free: 5/min, Light: 60, Standard: 120, Premium: 500)
• Retry with backoff — automatic retry for 429/5xx errors
• Pagination — transparent multi-page fetching
• Plan-aware — all tools registered regardless of plan; graceful error messages on restriction

Requirements

• Python 3.10+
• J-Quants API key (Free plan or above)

Installation

# Using uv (recommended)
uv pip install jquants-mcp

# Using pip
pip install jquants-mcp

From source

git clone https://github.com/shigechika/jquants-mcp.git
cd jquants-mcp
uv sync --dev

Configuration

Settings are loaded with the following priority (later wins):

1. ~/.jquants-api/jquants-api.toml — API key only (J-Quants official config)
2. ~/.config/jquants-mcp/config.ini (user global)
3. ./config.ini (current directory)
4. Environment variables (from MCP client or shell)

API Key (zero-config)

If you already use jquants-api-client, your API key is automatically read from ~/.jquants-api/jquants-api.toml. No extra configuration needed.

API Key via browser login

jquants-mcp login

Opens a browser to J-Quants (AWS Cognito, PKCE flow), and on success writes the API key to ~/.config/jquants-mcp/config.ini (mode 0600). Same auth backend as the official jquants-cli. Use jquants-mcp logout to clear the saved key.

config.ini

MCP-specific settings (cache, client behavior):

[jquants]
# cache_dir = ~/.cache/jquants-mcp
# base_url = https://api.jquants.com/v2

[client]
# max_retries = 5
# retry_base_delay = 1.0
# max_pages = 10

[server]
# encryption_key = <random-secret>   # enables per-user API key storage (multi-user mode)
# allowed_emails = [email protected],[email protected]

Environment Variables

* API key is auto-detected from ~/.jquants-api/jquants-api.toml. Set JQUANTS_API_KEY only to override.

Environment variables override both config.ini and jquants-api.toml. This allows MCP clients (Claude Desktop, Claude Code, Codex, …) to pass settings via their env block while keeping defaults elsewhere.

macOS launchd note

If you run jquants-mcp as a macOS LaunchAgent and the API key lives in ~/.jquants-api/jquants-api.toml, the server may silently hang during startup on macOS 26 or later. The TCC sandbox applied to launchd-spawned processes blocks open() on some dotfiles under $HOME (mode 600), and the process never finishes starting.

Workaround: copy the toml outside the sandboxed home hierarchy and point the server at it via JQUANTS_API_TOML_PATH:

sudo mkdir -p /usr/local/etc/jquants-mcp
sudo cp ~/.jquants-api/jquants-api.toml /usr/local/etc/jquants-mcp/jquants-api.toml
sudo chown "$USER":staff /usr/local/etc/jquants-mcp/jquants-api.toml
sudo chmod 600 /usr/local/etc/jquants-mcp/jquants-api.toml

Then add the following to your LaunchAgent plist's EnvironmentVariables dict:

<key>JQUANTS_API_TOML_PATH</key>
<string>/usr/local/etc/jquants-mcp/jquants-api.toml</string>

Alternatives: set JQUANTS_API_KEY directly in the plist (simpler but puts the key in a plist file that Time Machine / iCloud may back up), or put api_key = directly in ~/.config/jquants-mcp/config.ini (if that path is not sandbox-blocked on your macOS version).

Linux/systemd and other init systems are not affected.

Authentication

The server speaks stdio only and binds no network socket, so it performs no authentication of its own — there is no listener to authenticate against. Access control is a property of whoever starts the process:

Gateway identity

For remote or multi-user deployments, put a gateway in front of the stdio server. mcp-stdio serve is the gateway both production deployments use: it accepts MCP over HTTP, authenticates the caller, and spawns one jquants-mcp child process per session.

The gateway passes the authenticated identity down to that child as the JQUANTS_MCP_USER environment variable (mcp-stdio serve --trusted-user-header X-Forwarded-Email --user-env JQUANTS_MCP_USER). One child process serves exactly one principal for its whole lifetime, and the server reads JQUANTS_MCP_USER as that principal's verified email — it applies the JQUANTS_ALLOWED_EMAILS allowlist and per-user rate limits to it, and looks up that user's stored API key.

Trust model: JQUANTS_MCP_USER is trusted as-is; the server does no verification of its own. Whoever can set the child process's environment is that user, so the variable must be injected by the gateway and never by a user-supplied value.

When JQUANTS_MCP_USER is absent — the normal local case — the server runs single-user against the globally configured API key.

A worked example of the whole shape (auth proxy in front, gateway, stdio child, cache download) is scripts/entrypoint-stdio.sh, the entrypoint of the Cloud Run deployment described below. There is no other documented remote-access story: the server itself has no transport, TLS, or token options.

Multi-user Mode

When the server receives a gateway identity (JQUANTS_MCP_USER, see Authentication) and MCP_ENCRYPTION_KEY is configured, it operates in multi-user mode: each user stores their own J-Quants API key on the server, and all data tools use that key automatically. All users share the read cache; each user gets an independent J-Quants client with isolated rate limiting and their own plan's date-range window.

User flow

sequenceDiagram
    participant U as User
    participant C as MCP client
    participant G as Gateway (mcp-stdio serve)
    participant S as jquants-mcp (stdio child)
    participant J as J-Quants API
    U->>C: Connect
    C->>G: Authenticate (gateway's own auth layer)
    G->>S: Spawn child with JQUANTS_MCP_USER = verified email
    U->>C: "Register my J-Quants API key: <key>"
    C->>G: register_api_key(api_key="<key>")
    G->>S: forward over stdio
    S->>J: Probe plan-specific endpoints (auto-detect)
    J-->>S: Detected plan
    S->>S: Encrypt & store key + plan (AES-256-GCM)
    S-->>C: {"status": "ok", "plan": "<detected>"}
    U->>C: "Get TOPIX daily prices"
    C->>S: get_indices_bars_daily_topix(...)
    S->>J: API call with user's key
    J-->>S: Data
    S-->>C: Result

Tools for multi-user mode

Both tools return an explanatory error instead of failing silently when either requirement is missing.

Registering a key (tell your AI agent):

"Register my J-Quants API key: <your-api-key>"

The agent calls register_api_key(api_key="..."). The server probes plan-specific endpoints with the key to auto-detect the plan (free / light / standard / premium) and stores it alongside the encrypted key — no manual selection needed. Subsequent tool calls use the detected plan for rate limiting and date-range restrictions.

This is the only way to register a key on a multi-user deployment; the server exposes no web form and no HTTP endpoint of any kind.

Security

• API keys are encrypted with AES-256-GCM (authenticated encryption — integrity-protected)
• The encryption key is derived via PBKDF2-HMAC-SHA256 (200,000 iterations) from MCP_ENCRYPTION_KEY, with a random 16-byte salt per encryption
• Each ciphertext uses a unique random 12-byte nonce — encrypting the same key twice produces different ciphertext
• Tampered or truncated ciphertexts are rejected before decryption

Single-user fallback

Usage

Claude Code (plugin)

This repository doubles as a single-plugin marketplace, so Claude Code can install the server for you:

/plugin marketplace add shigechika/jquants-mcp
/plugin install jquants-mcp@jquants-mcp

The plugin launches uvx, so it must be on the PATH of the process that runs Claude Code — a login shell usually has it, but a GUI-launched app may not; install uv system-wide if the plugin fails to start.

The plugin launches uvx jquants-mcp and reads the same environment variables described in Configuration. JQUANTS_API_KEY is deliberately left out of the plugin's own .mcp.json so the key you already have — from jquants-mcp login (~/.config/jquants-mcp/config.ini) or the J-Quants official ~/.jquants-api/jquants-api.toml — is still found; export it yourself before starting Claude Code only if you want to override both of those.

Claude Code (manual)

Register the MCP server with claude mcp add:

claude mcp add jquants-mcp -- jquants-mcp

Or if installed from source:

claude mcp add jquants-mcp \
  -- /path/to/jquants-mcp/.venv/bin/jquants-mcp

The --scope (-s) option controls where the configuration is stored:

API key is auto-detected from ~/.jquants-api/jquants-api.toml. Set --env JQUANTS_API_KEY=... only to override.

AI Agent Skills

Install the operational guidance Skill into your Claude Code project:

npx skills add shigechika/jquants-mcp

This adds skills/jquants-mcp-usage/SKILL.md to your project, giving Claude Code the daily workflows (one-call briefings, value screening, single-stock deep dives) plus practical tips on cache tiers, plan-based date limits, screener patterns, and safe cache management — without touching the tool definitions.

Claude Desktop

Add to Claude Desktop config file:

{
  "mcpServers": {
    "jquants-mcp": {
      "command": "/path/to/jquants-mcp/.venv/bin/jquants-mcp"
    }
  }
}

The server auto-detects the plan from your API key on startup — no need to set it manually. Add an env block only if you want to override the detection or point to a different API key.

Note: Claude Desktop has a limited PATH (/usr/local/bin, /usr/bin, etc.), so you must specify the full path to the executable.

Restart Claude Desktop after editing.

Codex

codex mcp add jquants-mcp -- jquants-mcp

API key is auto-detected from ~/.jquants-api/jquants-api.toml. Set --env JQUANTS_API_KEY=... only to override. Any other MCP client that launches local stdio servers uses the same command.

Standalone (stdio)

jquants-mcp

Remote access

The server has no network transport of its own — no --transport, no --host/--port, no TLS options. To reach it from another machine, front the stdio server with a gateway that terminates authentication and spawns jquants-mcp per session. With mcp-stdio serve that looks like:

mcp-stdio serve --host 127.0.0.1 --port 8081 --path /mcp \
  --enable-oauth --public-url https://mcp.example.com \
  --trusted-user-header X-Forwarded-Email --user-env JQUANTS_MCP_USER \
  -- jquants-mcp

--trusted-user-header requires an authenticating proxy in front — not merely a TLS-terminating one. The gateway trusts that header without verifying it. A plain TLS proxy passes client headers through, so anyone could send X-Forwarded-Email: [email protected] and be served as that user, reaching their stored API key. The proxy must authenticate the caller and overwrite the header with the verified identity, discarding whatever the client sent. oauth2-proxy does this (its skip_auth_strip_headers default strips client-supplied headers on pass-through routes). Bind the gateway to loopback, as above, so it is reachable only via that proxy.

Omit --trusted-user-header and --user-env if you have no such proxy: the server then runs single-user against the configured API key, which is safe.

scripts/entrypoint-stdio.sh is the worked example. It is the entrypoint of the Cloud Run deployment described below, and shows the gateway running behind an oauth2-proxy sidecar together with the startup cache download.

That is the whole remote-access story. How clients connect depends only on the gateway you chose; the sections below cover the two shapes used with mcp-stdio serve.

Stdio clients via mcp-stdio (Claude Code, Claude Desktop, Codex, …)

Claude Desktop does not speak Streamable HTTP directly, and claude mcp add --transport http --header "Authorization: Bearer ..." drops the header during health checks (claude-code#28293). The mcp-stdio client bridges stdio to the remote endpoint in both cases, and the same bridge works for any other client that launches stdio servers:

pip install mcp-stdio  # or: uvx mcp-stdio

# Claude Code — --oauth drives the gateway's OAuth flow in your browser
claude mcp add jquants-mcp -- mcp-stdio --oauth https://mcp.example.com/mcp

# Codex
codex mcp add jquants-mcp -- mcp-stdio --oauth https://mcp.example.com/mcp

{
  "mcpServers": {
    "jquants-mcp": {
      "command": "mcp-stdio",
      "args": ["--oauth", "https://mcp.example.com/mcp"]
    }
  }
}

Restart Claude Desktop after editing. Use --bearer-token <TOKEN> instead of --oauth if your gateway authenticates with a static token.

Web clients (claude.ai, Meta Muse)

Browser-based clients connect to the gateway's OAuth endpoint natively. They register through Dynamic Client Registration with a fixed HTTPS callback, so the gateway must allow each client's callback explicitly (matched byte-for-byte):

mcp-stdio serve ... --enable-oauth \
  --allow-redirect-uri https://claude.ai/api/mcp/auth_callback \
  --allow-redirect-uri https://agent.meta.ai/api/hatch/oauth/callback \
  -- jquants-mcp

The first is claude.ai (custom connectors, also used by the Claude mobile apps), the second Meta Muse. Then add https://mcp.example.com/mcp as a connector in the client and complete the login in the browser.

Claude Desktop Connectors

When the gateway speaks OAuth 2.1 over HTTPS (mcp-stdio serve --enable-oauth), Claude Desktop's Connectors panel can connect natively — click Connect and complete the login in the browser; the token is stored and reused silently.

{
  "mcpServers": {
    "jquants-mcp": {
      "type": "http",
      "url": "https://mcp.example.com/mcp"
    }
  }
}

Per-user API keys additionally require MCP_ENCRYPTION_KEY on the server — see Multi-user Mode. If "type": "http" is not yet available in your Claude Desktop version, use the mcp-stdio bridge above instead.

Available Tools

Equities (8 tools)

Financials (4 tools)

Indices (2 tools)

Derivatives (3 tools)

Markets (6 tools)

Bulk Download (2 tools)

Market Overview & Valuation (11 tools)

Cross-sectional cache-only tools that scan all listed equities. No extra API calls, useful for "what's the overall market doing today?" and sector valuation queries.

Screener (10 tools)

Offline tools that compute signals directly from the SQLite cache. No extra API calls, pure Python, no numpy/pandas. Intended for AI-assisted stock screening without hitting rate limits.

Single Stock Briefing (1 tool)

Cache-only tool that assembles a one-page snapshot for a single stock from cached data. No extra API calls.

Technical Indicators (1 tool)

Pure-Python SMA / Bollinger Bands / RSI computation over the cached daily bars. No extra API call for codes already in cache; falls back to the J-Quants API on a cache miss and stores the result.

RSI in charts: RSI sub-panel is not yet available. Use get_technical_indicators for numeric RSI values.

Charts (2 tools)

Both tools return JSON for React artifact / Plotly rendering (no optional dependencies).

Indicator options for get_candlestick_data:

• Indicators: volume, sma5, sma20, sma25, sma60, sma75, sma200, bb20 (20-day Bollinger band; expands to bb20_upper / bb20_mid / bb20_lower)
• Adjusted prices: split-adjusted by default (adjusted=True); set False for raw OHLC

Utility (5 tools)

Caching

The server uses a two-tier SQLite cache:

• Tier 1 (Row-level): Time-series data cached by date and code. Supports incremental fetching and stock split detection via AdjFactor comparison.
  • equities_bars_daily, equities_master, fins_summary, indices_bars_daily_topix, investor_types, markets_margin_interest, markets_margin_alert, markets_short_ratio, markets_breakdown, markets_calendar
• Tier 2 (Response-level): Full API responses cached with a TTL (6h / 24h / 7d by endpoint). The daily /markets/* answers instead expire at the endpoint's next publication time from the J-Quants data update schedule, and are re-checked shortly after it until the day's delivery shows up (cache/publication.py).

Cache is stored at ~/.cache/jquants-mcp/cache.db by default.

Expected disk usage after a full historical fetch (approximate; varies by market data availability):

Bulk Data Import

The scripts/bulk_fetch_all.py script downloads all available bulk CSV data from the J-Quants Bulk API and imports it into the SQLite cache. This is the fastest way to populate the local cache with historical data.

# Fetch all available data for your plan
uv run python scripts/bulk_fetch_all.py

# Fetch specific endpoints only
uv run python scripts/bulk_fetch_all.py --endpoints fins_summary topix margin_interest

# Dry run — show file list and sizes without downloading
uv run python scripts/bulk_fetch_all.py --dry-run

The script respects the plan-based rate limit (e.g. 60 req/min for Light) and retries on 429 errors. A full historical fetch takes roughly 1 hour; use health_check to monitor progress.

CSV Import

The CSV sideload script (import_csv_to_cache.py) is maintained by the publisher pipeline that feeds this cache. If you are building your own pipeline, implement sideloading by inserting directly into the equities_bars_daily / equities_master tables following the schema defined in src/jquants_mcp/cache/schema.py.

Daily Fetch

scripts/daily_fetch.py fetches additional J-Quants data via jquantsapi.ClientV2 and inserts it directly into the SQLite cache. Designed to be called from an external daily pipeline (e.g. a cron job or shell script).

jquants-api-client is not a core dependency (the MCP server itself never imports it) — install it with the fetch extra: uv sync --extra fetch. Without it, a routine uv sync on a self-hosted deployment will silently remove a manually-installed copy.

The script reads the plan from ~/.config/jquants-mcp/config.ini (or JQUANTS_PLAN env var) and automatically determines which endpoints to fetch:

# Fetch all endpoints available for your plan
python3 scripts/daily_fetch.py

# Fetch specific endpoints only
python3 scripts/daily_fetch.py --topix --investor-types

# Fetch trading calendar
python3 scripts/daily_fetch.py --calendar

# Backfill historical Markets data (past N days)
python3 scripts/daily_fetch.py --backfill 90

# Use a custom cache DB path
python3 scripts/daily_fetch.py --db /path/to/cache.db

Permission errors (403) are handled gracefully — the script logs the error and continues to the next endpoint without crashing.

Cache Health Check

scripts/verify_cache_completeness.py audits the local cache and reports which tables are up-to-date, stale, or missing for the current plan.

# Quick freshness check (text output)
uv run python scripts/verify_cache_completeness.py

# Machine-readable JSON (for CI / monitoring)
uv run python scripts/verify_cache_completeness.py --output json

# Detect date-level gaps (days where only a fraction of stocks were fetched)
uv run python scripts/verify_cache_completeness.py --check-gaps

# Show what --auto-fix would repair, without making API calls
uv run python scripts/verify_cache_completeness.py --check-gaps --auto-fix --dry-run

# Re-fetch gap days automatically
uv run python scripts/verify_cache_completeness.py --check-gaps --auto-fix

Exit codes: 0 = all tables healthy, 1 = stale or missing tables, 2 = fatal (DB unreadable).

The plan is auto-detected from your API key (same probe as daily_fetch.py); pass --plan <plan> or set JQUANTS_PLAN to override (skips the probe).

Useful before a plan downgrade to confirm all currently-covered data has been fetched, and as a periodic check to catch silent fetch failures early.

Cloud Run Deployment

This server can be deployed to Google Cloud Run. Because the server is stdio-only, the service is two containers: an oauth2-proxy sidecar handles Google login, and the app container runs scripts/entrypoint-stdio.sh — mcp-stdio serve fronting a per-session jquants-mcp child process (see Authentication).

State is split across two managed stores:

• cache.db — published to a GCS bucket by the self-hosted server and downloaded to /tmp (tmpfs) on every cold start. Cloud Run reads it but never writes back.
• users — per-user encrypted J-Quants API keys, stored in Firestore (Native mode). Strongly consistent and multi-writer safe, so no SQLite write conflicts. The gateway keeps its own OAuth token state in a separate Firestore collection.

Details: see GCS and Firestore integration below.

For a fork-and-deploy walkthrough (WIF, OAuth client, custom domain, Claude mobile setup, allowlist), see docs/deploy/gcp.md. The sections below summarise the moving parts; the deploy guide is the canonical step-by-step.

Prerequisites

• Google Cloud SDK
• A GCS bucket holding a read-only snapshot of cache.db (updated out-of-band by the self-hosted server)
• Firestore in Native mode enabled on the project (stores per-user API keys, and the gateway's OAuth token state)
• A service account with:
  • roles/storage.objectViewer on the GCS bucket (read-only access to cache.db)
  • roles/datastore.user on the project (Firestore read/write)
  • roles/secretmanager.secretAccessor if using Secret Manager for API keys

Create a GCS bucket

gcloud storage buckets create gs://YOUR_BUCKET \
  --location asia-northeast1

Enable Firestore

gcloud firestore databases create \
  --location=us-west1 \
  --type=firestore-native

Deploy

The recommended path is to fork the repository and rely on the GitHub Actions CD workflow at .github/workflows/cd.yml. It builds the image with Cloud Build and then updates only the app container's image (gcloud run services update --container app --image …); the sidecar, scaling, CPU, env vars and secrets are set once by hand and deliberately never touched by CD.

The two-container service itself is created once, out of band, with the app container started as --command /app/scripts/entrypoint-stdio.sh. docs/deploy/gcp.md is the canonical step-by-step for that initial setup.

Memory and scaling sizing notes are in Memory requirements below.

Environment variables

Firestore uses Application Default Credentials from the Cloud Run service account. The oauth2-proxy sidecar carries its own configuration (Google OAuth client, cookie secret) — see docs/deploy/gcp.md.

GCS and Firestore integration

Cloud Run deployments depend on two managed stores, not an in-container SQLite set:

cache.db is owned by a self-hosted publisher (a cron / scheduled task running scripts/daily_fetch.py or scripts/bulk_fetch_all.py + scripts/gcs_export_cache.py) that pushes a fresh snapshot to GCS on each run. Cloud Run never writes back to GCS.

Startup flow

sequenceDiagram
    participant E as entrypoint-stdio.sh
    participant G as GCS
    participant M as mcp-stdio serve

    E->>G: download cache.db.zst to /tmp (synchronous)
    Note right of E: runs in the startup window,<br/>where CPU is fully allocated
    G-->>E: ~1.2 GiB compressed
    Note right of E: stream-decompressed to ~3 GiB;<br/>falls back to uncompressed cache.db
    E->>M: start gateway (cache.db already present)
    activate M
    Note right of M: each session spawns a stdio child<br/>serving from the Tier 1 cache<br/>(live J-Quants API only if the<br/>download was skipped/failed)
    deactivate M

Notes:

• cache.db is downloaded synchronously during container startup, before the gateway starts accepting sessions. It is published zstd-compressed as cache.db.zst (~1.2 GiB on the wire, stream-decompressed to ~3 GiB) because the Cloud Run instance's GCS read bandwidth (~60 MB/s) is the bottleneck; the downloader falls back to the uncompressed cache.db when .zst is absent. It runs in the container-startup window, where CPU is fully allocated (plus --cpu-boost): under Cloud Run's request-based CPU allocation a download started after the server is ready can be throttled to ~0 between requests and never finish. The trade-off is a longer cold start — the first request after a scale-to-zero waits for the download.
• If the download fails, startup continues and the server serves via the live J-Quants API (slower, counts against rate limits). cache_status then returns a minimal payload (db_path + plan only) until a cache is loaded.
• Firestore is strongly consistent, so per-user data survives instance recycling and is safe against concurrent writers. Sessions, however, are held in-instance by the gateway (one child process per session), so the production service runs with max-instances=1; scaling out is a gateway-level concern, not a storage one.

Daily cache refresh

After startup, cache.db is refreshed daily by the publisher. How that update reaches a running server depends on the deployment.

Cloud Run — instance recycling

There is no in-container refresh mechanism. With min-instances=0 every cold start downloads a current cache.db, so the only window in which a running instance can hold a stale copy is one that stays warm across the publisher's export. In that window, days not yet cached fall through to the live J-Quants API (correct, just slower), while corrections to already-cached rows do stay stale — the cache-vs-API decision is presence-based and the row-level tier applies no TTL. Measured instance lifetimes under min-instances=0 are 15–26 minutes, so the exposure is bounded by recycling. A push-based reload endpoint existed until v1.0.0; it was removed after the design was evaluated and rejected as not worth the moving parts (#584).

Local process

The publisher and the server share a filesystem, so an updated cache.db is visible to the next query with no signal required. Behind a gateway such as mcp-stdio serve, each session spawns a fresh child process that opens the file as it stands at that moment.

Troubleshooting

Permission error on startup (403 Forbidden or storage.objects.get denied):

gcloud storage buckets get-iam-policy gs://YOUR_BUCKET \
  --format="table(bindings.role, bindings.members)"

The service account needs roles/storage.objectViewer on the bucket — see IAM setup.

Firestore permission errors:

gcloud projects get-iam-policy "${PROJECT_ID}" \
  --flatten="bindings[].members" \
  --filter="bindings.members:serviceAccount:jquants-mcp@*"

The service account needs roles/datastore.user on the project.

cache_status returns only db_path and plan (no row counts):

The startup download did not complete, or the integrity prewarm is still running. Check the container log for cache.db download complete (emitted by entrypoint-stdio.sh after the synchronous download); its absence means the download was skipped or failed and the server is on the live-API fallback.

cache.db not found in GCS on first deploy:

There is no "empty cache" fallback mode beyond API fallback — the server will keep serving requests directly from the J-Quants API. Upload a cache.db snapshot from your self-hosted server to GCS to enable Tier 1 caching (see Initial cache.db upload).

IAM setup

SA="jquants-mcp@${PROJECT_ID}.iam.gserviceaccount.com"

# Create service account
gcloud iam service-accounts create jquants-mcp \
  --display-name "jquants-mcp Cloud Run SA"

# Read-only access to the cache.db snapshot in GCS
gcloud storage buckets add-iam-policy-binding gs://YOUR_BUCKET \
  --member "serviceAccount:${SA}" \
  --role "roles/storage.objectViewer"

# Firestore access for the users collection and the gateway's token store
gcloud projects add-iam-policy-binding "${PROJECT_ID}" \
  --member "serviceAccount:${SA}" \
  --role "roles/datastore.user"

# Secret Manager access (if using Secret Manager for JQUANTS_API_KEY etc.)
gcloud projects add-iam-policy-binding "${PROJECT_ID}" \
  --member "serviceAccount:${SA}" \
  --role "roles/secretmanager.secretAccessor"

Note: if the self-hosted server that publishes cache.db uses a different service account, only that account needs write access to the bucket. The Cloud Run service account remains viewer-only.

Initial cache.db upload

Cloud Run reads cache.db as a read-only snapshot. Publish a snapshot from your self-hosted server (which has been warming the cache) before the first deploy:

gcloud storage cp ~/.cache/jquants-mcp/cache.db \
  gs://YOUR_BUCKET/jquants-mcp/cache.db \
  --no-gzip-in-flight

Important: disable parallel composite uploads (the default for large files). They corrupt SQLite files because the reassembled object contains byte ranges that do not form a valid database page layout. On the publishing host, set:

gcloud config set storage/parallel_composite_upload_enabled False

No manual Firestore setup is required — collections are created on first write.

Memory requirements

Cloud Run materializes cache.db into /tmp (a tmpfs, i.e. RAM). The memory limit therefore must cover:

• cache.db size (currently ~3 GiB)
• Python runtime + mcp SDK + sqlite + httpx overhead (~300 MiB)
• Request-time JSON serialization headroom

Current production sizing is --memory 8Gi --cpu 2, with min-instances=0 (scale to zero), max-instances=1, and CPU always allocated. Those three billing-relevant settings are never passed by the CD workflow; they are set once by hand and asserted before and after every deploy by .github/workflows/scripts/assert-jquants-billing-settings.sh, so an out-of-band change fails the next deploy instead of persisting silently. Scale-to-zero is load-bearing for correctness as well as cost — every cold start re-downloads a current cache.db (see Daily cache refresh). Cloud Run gen2 is required for memory allocations above 4 Gi, and >4 GiB also forces ≥2 vCPU (8 GiB is the ceiling for 2 vCPU).

Memory is 8 GiB because a cache reload briefly holds ~2× cache.db in /tmp (which is tmpfs, i.e. RAM): the new snapshot downloads to a temp file while the current cache.db is still mapped, then atomically replaces it. At ~3 GiB per snapshot that peak (~6 GiB) plus the Python/SQLite RSS exceeds a 6 GiB limit and tmpfs writes fail with SIGBUS (observed as Container terminated on signal 7), so the limit is 8 GiB. If cache.db grows materially, raise the limit further (and keep ≥2 vCPU; >8 GiB needs ≥4 vCPU).

Operations

For production incidents on the Cloud Run deployment, see the runbooks:

• OOM / memory pressure
• 5xx spike
• Firestore outage / quota
• cache.db missing / download failed
• OAuth loop / persistent 401
• Firestore restore
• Secrets rotation

Alert policies that trigger these are in ops/alerts/; each policy's documentation links back to the matching runbook.

The disaster recovery posture documents the current single-region deployment, RTO/RPO expectations, and the (undrilled) standby-region procedure.

Service-level objectives — availability and latency targets with an error-budget policy — are in docs/slo.md.

Development

# Install dev dependencies
uv sync --dev

# Run tests
uv run pytest -v

# Lint
uv run ruff check src/ tests/

# Format
uv run ruff format src/ tests/

Disclaimer

This software (jquants-mcp) is a technical tool for retrieving Japanese stock data from the J-Quants API v2 for use with Claude and other MCP clients. It is intended to provide reference information for your own investment research, and:

• This software and its output do not constitute investment advice or recommendations.
• We make no warranty regarding the accuracy, completeness, or timeliness of the information provided.
• Investment decisions are made at your own risk and responsibility.
• Past performance does not guarantee future results.
• The author is not registered as a financial instruments business operator under Japanese law.
• Use is subject to the terms and conditions of J-Quants, the underlying data provider.
• The author disclaims all liability for any damages arising from the use of this software.

License

MIT