mcp-cache

A reusable SQLite-backed cache for MCP (Model Context Protocol) servers. Zero dependencies — pure Python stdlib.

Two caching strategies:

• TTL cache — store any JSON-serializable value for N seconds
• Time series cache — store date-keyed observations and automatically fetch only missing date ranges

Installation

pip install mcp-cache

Or from source:

git clone https://github.com/cmaurer/mcp-cache
cd mcp-cache
pip install -e .

Requires Python 3.10+.

Quick start

TTL cache

from mcp_cache import MCPCache

cache = MCPCache("~/.cache/myserver.db", default_ttl=300)

async def get_quote(symbol: str) -> dict:
    return await cache.get_or_fetch(
        key=f"quote:{symbol}",
        fetch_fn=lambda: api.fetch_quote(symbol),
        ttl=60,  # override default; omit to use default_ttl
    )

fetch_fn is called only on a cache miss or after the TTL expires. The result is stored as JSON and returned on subsequent calls within the TTL window.

Time series cache

observations = await cache.get_timeseries(
    series_id="T10YIE",
    start_date="2020-01-01",
    end_date="2024-12-31",
    fetch_fn=lambda s, e: fred_api.get_series("T10YIE", s, e),
)
# Returns list of {"date": "YYYY-MM-DD", "value": float} dicts, newest first

On the first call the full range is fetched. On subsequent calls only gaps are fetched — requesting a wider range re-uses the already-cached portion and fetches only the missing edges.

Custom key names are supported for APIs that don't use "date" / "value":

observations = await cache.get_timeseries(
    series_id="prices",
    start_date="2024-01-01",
    end_date="2024-06-30",
    fetch_fn=my_fetch,
    date_key="timestamp",
    value_key="close",
)

API reference

MCPCache(db_path, default_ttl)

Parameter	Type	Default	Description
db_path	str | Path	"mcp_cache.db"	Path to SQLite file. ~ is expanded. Parent dirs are created automatically.
default_ttl	int	300	Default TTL in seconds for get_or_fetch.

---

TTL cache methods

await cache.get_or_fetch(key, fetch_fn, ttl=None)

Return the cached value for key if it exists and is still fresh. Otherwise call fetch_fn(), store the result, and return it.

• key — cache key string
• fetch_fn — async callable that returns a JSON-serializable value
• ttl — per-call TTL override in seconds; uses default_ttl if omitted

await cache.invalidate(key)

Remove a specific entry from the TTL cache. No-op if the key does not exist.

await cache.clear_expired()

Delete all expired TTL entries. Returns the number of entries removed.

await cache.list_keys(prefix=None, include_expired=False)

Return TTL cache keys as a sorted list of strings.

• prefix — only return keys starting with this string; None (default) returns all keys
• include_expired — include expired entries; defaults to False (fresh keys only)

---

Time series methods

await cache.get_timeseries(series_id, start_date, end_date, fetch_fn, date_key="date", value_key="value")

Return cached observations for series_id in [start_date, end_date].

• series_id — identifier for the series
• start_date, end_date — date objects or ISO strings ("YYYY-MM-DD")
• fetch_fn(start, end) — async callable receiving ISO date strings; must return a list of dicts containing date_key and optionally value_key
• date_key, value_key — key names in the returned dicts (default "date", "value")

Returns a list of {date_key: str, value_key: float | None} dicts sorted newest first. None values are preserved — they represent real data points (e.g. non-trading days).

await cache.invalidate_series(series_id)

Remove all cached observations and range records for a series.

await cache.list_series(prefix=None)

Return distinct time series IDs as a sorted list of strings.

• prefix — only return series IDs starting with this string; None (default) returns all series

---

Diagnostics

await cache.stats()

Returns a dict with cache counts:

{
    "ttl_cache":  {"total": 12, "fresh": 10, "expired": 2},
    "timeseries": {"series": 3, "observations": 1500, "fetched_ranges": 6},
    "db_path":    "/home/user/.cache/myserver.db",
}

Using with an MCP server

from mcp_cache import MCPCache

_cache = MCPCache("~/.cache/myserver.db")

@server.tool()
async def get_price_history(symbol: str, start: str, end: str) -> list[dict]:
    return await _cache.get_timeseries(
        series_id=symbol,
        start_date=start,
        end_date=end,
        fetch_fn=lambda s, e: data_provider.fetch(symbol, s, e),
    )

Using with Claude Desktop

Install the package:

pip install mcp-cache

Add the server to ~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "cache": {
      "command": "mcp-cache"
    }
  }
}

Restart Claude Desktop. The following tools will be available:

Tool	Description
cache_get(key)	Return a cached value, or null if missing/expired
cache_set(key, value, ttl)	Store a JSON value for ttl seconds (default 300)
cache_invalidate(key)	Remove a specific entry
cache_clear_expired()	Delete all expired entries
cache_list_keys(prefix, include_expired)	List TTL cache keys, optionally filtered by prefix
timeseries_store(series_id, observations, range_start, range_end)	Store date-keyed observations
timeseries_get(series_id, start_date, end_date)	Query cached observations
timeseries_invalidate(series_id)	Clear all data for a series
timeseries_list_series(prefix)	List time series IDs, optionally filtered by prefix
cache_stats()	Return entry counts and db path

The cache database is stored at ~/.cache/mcp_cache.db.

Development

pip install -e ".[dev]"
pytest

License

MIT