The MCP server for AI assistants
Connect Claude or any MCP client to your plan with 54 read and compute tools and 21 write tools that stay off by default. Edits are always proposals.
The Model Context Protocol is the standard way for AI assistants to use tools. Glideworth’s MCP server lets an assistant read your plans, run simulations and propose changes, under rules you set. It is a thin adapter over the same services as the REST API, so every tool has a REST twin and behaves identically.
What an assistant gets
There are 75 tools. 54 read and compute, and all of them are marked read-only. 21 write, and they exist only when you switch writes on at the server and the connecting key has write scope. A client cannot call a tool it cannot see.
- Read: households, accounts, plans, plan history, diffs, progress snapshots, check-in history, notifications, exchange rates, the residency timeline and cross-border rule tables, and dry-run previews of CSV and bank-file imports.
- Compute: a plan summary (the best first call), projections, Monte Carlo and historical backtests, plan comparison, explanations, the assumption audit, sensitivity, every solver, guardrails, tax breakdowns, stay-or-move, and previews of what-ifs.
- Write: propose, apply and discard change sets; create branches; apply imports; update balances; record snapshots; run check-ins; create and revoke share links; set exchange rates and a person’s cross-border record.
- Resources and prompts: reference documents (units, glossary, change sets, common edit paths) and prompts
such as
review_my_plan,why_might_my_plan_fail,how_much_can_i_spendandcan_i_retire_earlier.
Tools return compact JSON with money in today’s money unless asked otherwise, and every error says how to recover: a code, a message and a hint about which tool to call next.
The safety model
- Read-only by default. The write switch is off until you turn it on.
- Scoped keys.
readfor analysis,simulatefor compute only,writeto propose and apply. A viewer’s connection never has write tools, whatever the key says. - Edits are proposals.
propose_change_setreturns a readable diff and a before-and-after result and applies nothing.apply_change_setworks only against the plan version the proposal was written for, or it is refused as a conflict. - No delete tools. Anything destructive has to be a change set, even reverting to an earlier version.
- An audit trail. Every apply is recorded in plan history under the assistant’s identity, and writes appear in the audit log under the person whose key or consent it was.
- Limits that hold under manipulation. An assistant that has been tricked by text it read can still do only what its key allows, which is why least-privilege keys matter.
A worked illustration
You ask, “What if I retire three years earlier? Save it as a branch.” Illustrative figures throughout. The
assistant calls get_plan_summary, then preview_whatif with the early-retirement template and years: 3
(nothing stored), then create_branch_from_template (the parent is untouched), run_monte_carlo on both plans and
compare_plans with an explanation. It reports: “Retiring at 57 instead of 60 lowers the chance of success from
86% to 74%. It is saved as a branch called Retire at 57.” You open the branch in the app and decide. The
API and MCP page walks through a longer session, including a change set.
Connecting
Over HTTP against a running server, or over stdio with no server running, with the commands shown on the API and MCP page. Remote assistants that connect by URL sign in with OAuth 2.1 (PKCE, rotating refresh tokens, scopes capped by your role) instead of a pasted key. Bearer keys keep working.
What it does not do
- Glideworth does not run an AI on your data. The server has no model in it. Your assistant is software you choose, and its provider sees what the tools return. If that matters, use a local model.
- An assistant is not a source of tax or investment advice. Tools return the model’s results; what the assistant tells you about them is its own.
- No hosted endpoint yet. A stable HTTPS address with OAuth is planned for the hosted service, which is not open yet. Today the server runs on your machine.
- Some tools need the writes switch, and a tool that touches an engine feature that is not built returns
NOT_IMPLEMENTEDwith the owner named. - It does not replace the app. Charts, the plan editor and some screens have no MCP equivalent.
Go further
- API and MCP for AI agents for the full session and connection snippets.
- Branches and change sets, the mechanism behind safe edits.
- Security and privacy for what a connected assistant can see.
See where your plan lands.
Free to self-host today, with every feature. The hosted service opens later, at $9 a month or $90 a year.