MCP server overview and safety model
What the Glideworth MCP server is, how read-only access and change sets keep AI assistants safe, how writes are enabled, and the prompts and resources it offers.
The Model Context Protocol (MCP) lets an AI assistant such as Claude Desktop, Claude Code or Cursor call tools. Glideworth’s MCP server lets an assistant read your plans, run simulations and solvers, and propose edits that you approve. It is a thin adapter over the same services as the REST API, so every tool has a REST twin.
The server has 75 tools: 54 that read and compute (all marked read-only) and 21 that write, which exist only when writes are enabled. See the tool list.
The safety model
- Read-only by default. Write tools register only when the server enables writes
(
GLIDEWORTH_MCP_ALLOW_WRITES=true) and the connecting key haswritescope. A client that cannot see a tool cannot call it. Aviewernever has write tools. - Plan edits are change sets or branches, never overwrites. The assistant proposes JSON Patch operations, gets a readable diff and a before and after outcome, shows you, and only then applies. A stale proposal is a conflict, not a silent merge. Every apply is recorded in plan history with the assistant’s identity.
- Household data is different by design.
update_account_balances,record_progress_snapshotandapply_csv_importwrite directly, because they record actuals, not assumptions, and are atomic.create_branchandcreate_branch_from_templateadd a what-if without touching the parent. - No delete tools. Even “revert to version N” is a proposal. The only undo-style tools are
revoke_share_linkandunset_fx_rate; neither deletes plans or household data. - Compact outputs. Tools return compact JSON with ids included, money rounded to whole units in today’s money, and a
detailofsummaryorfullwhere output can be large. - Errors say how to recover. A failure is an error result with
code,message,detailsand ahintsaying what to call next. A missing plan or household lists the valid ids. Unexpected failures say onlyinternal error. - Read means read. Read and compute tools never change household or plan data. Compute tools do write cache and job rows, which is why results are reproducible.
- Units are spelled out in the instructions the server sends: money in major units, rates as fractions,
realis today’s money.
Enabling write tools
Write tools need both:
- The server flag. On Docker or the server, set
GLIDEWORTH_MCP_ALLOW_WRITES=truein.env. For a stdio server, set it in the process environment. - A key with
writescope:pnpm keys:create -- --name claude-editor --scopes read,write. Over OAuth, the user must grantwrite, and their role must allow it.
Tool registration is decided per connection, so reconnect after changing either.
Prompts
Eight prompts guide common tasks. Each asks the assistant to use specific tools.
| Prompt | Arguments | What it asks |
|---|---|---|
review_my_plan |
planId? |
Outcome, Monte Carlo, audit and sensitivity, then the three most valuable changes. Changes nothing. |
why_might_my_plan_fail |
planId? |
Chance of success, then the failure analysis and explain_year on the earliest failure year. |
how_much_can_i_spend |
planId?, targetSuccessRate? |
solve_max_spending, quote the money per year, offer the edits. |
can_i_retire_earlier |
planId?, targetSuccessRate? |
solve_retirement_date, explain the trade-off, offer a branch. |
retire_earlier_as_branch |
planId?, years |
Preview, create a branch, run Monte Carlo on both, compare. |
check_my_assumptions |
planId? |
audit_plan, preview then propose the top fix, apply only after approval. |
what_if_spending |
planId?, change |
Turn “+500 a month on dining” into operations, preview, report before and after. |
monthly_checkin |
householdId? |
Collect balances, update them, compare with earlier snapshots and the plan. |
Resources
| URI | Content |
|---|---|
glideworth://schema/document |
JSON Schema of the whole household document (about 100 KB). |
glideworth://schema/plan |
JSON Schema of a plan (about 100 KB). |
glideworth://docs/edit-paths |
The dozen most common plan edits as ready-to-adapt operations. Read this before the big schemas. |
glideworth://docs/glossary |
Terms: household, plan, branch, milestone, flow, real versus nominal, change set. |
glideworth://docs/units |
Money, rates, bases, frequencies and patch addressing. |
glideworth://docs/change-sets |
The propose, review and apply workflow. |
glideworth://plans/{planId} |
The full plan document. |
glideworth://plans/{planId}/summary |
The compact plan. |
glideworth://households/{householdId}/document |
The whole workspace (large). |
What this does not do
It cannot delete plans or household data, overwrite a plan, or write when the flag is off. The assistant cannot see anything outside its key’s scope. The assistant’s explanations are only as good as the plan and the model.
Last checked against the code on .