The Model Context Protocol (MCP) is how Claude connects to external tools and data. The problem: most MCP servers require you to run code on your own machine — a Python process, a Node.js server, something that has to be running when you open Claude.
Stash is different. It's a hosted MCP server. You sign in with Google, get a connector URL, paste it into Claude, and you're done. Nothing to install. Nothing to keep running.
The typical MCP setup looks like this:
# Install the server npm install -g @modelcontextprotocol/server-filesystem # Start it before every Claude session npx mcp-server-filesystem /Users/you/Documents
If you forget to start it, your tools aren't there. If you switch machines, you set it up again. If you want to share access (say, the same memory across two computers), you need to synchronise files yourself.
A hosted MCP server solves all of this. The server runs in the cloud. Claude reaches it over HTTPS. Your data lives there, not in a process on your laptop.
Stash is a record store and context layer. It gives Claude four things:
| Factor | Self-hosted MCP | Stash (hosted) |
|---|---|---|
| Setup time | 15–60 min (install, configure, debug) | 30 seconds |
| Must be running | Yes — process must be live on your machine | No — always-on in the cloud |
| Multi-machine access | Complex (sync files, re-configure) | Same URL, same data, everywhere |
| Data durability | Wherever your files are | hosted on industry-secure servers, daily backups |
| Requires coding knowledge | Yes (Node/Python at minimum) | No |
| Cost | Free (if you count your time) | Free tier · Pro ~£8/mo (pricing may change) |
Every MCP response burns tokens from Claude's context window. If your server returns bulky JSON or full document bodies, that cost adds up fast.
Stash is designed to be token-light on purpose. Search results return record titles, tags, and a short excerpt — not the full body unless you ask for it. A 500-record search returns roughly 192 tokens (measured, preliminary, n=1). Fetching the equivalent data from a Notion integration returned ~410 tokens in the same test — about 2.1× more.
For context: if you call context() at the start of every conversation and
run 100 queries a month, Stash adds roughly $0.01–0.04 to your API bill depending on the model.
The context it saves you (not re-explaining who you are every session) is worth far more.
Store your role, current projects, working style, and preferences in Stash.
At the start of any conversation, call context(). Claude loads everything it needs
in one token-light call instead of you re-typing it or burning custom-instruction space.
You: context()
Claude: [loads your role, current projects, communication style]
Claude: Morning. You're in the middle of the Q3 roadmap review.
Your three live projects are X, Y, Z. What are we working on?
Keep a contacts collection in Stash. Add notes after meetings. Search before calls.
You: add to contacts — "Alice Chen, VP Eng at Acme, met 2026-06-01,
interested in API pricing, follow up in 2 weeks"
You: [two weeks later] search contacts for alice
Claude: Alice Chen — Acme, met June 1st. You were going to follow up on API pricing.
Notion is great for documents. It's not great as a Claude data source — Notion's API returns full page trees, and you hit a 25-result wall without pagination handling. Stash is a flat key-value store with full-text search. Migrate the rows you actually search, search them for a fraction of the token cost.
See the full Notion offload guide →
A dedicated per-account store databases, Google OAuth for authentication, HTTPS in transit. Each account's data is isolated — no shared namespace. Full security details at stashlite.com/security.
context() — you should get a response about your (empty) Stash.
Then: add to stash — "my first note"
That's it. No server to run, no config to manage. Stash handles the infrastructure; you get the MCP tools.