Rail
Budgets, propose-then-approve, and receipts for agents that buy for a human. Dry-run by default.
Installation
Ajoutez ceci à la configuration MCP de votre client (Claude Desktop : claude_desktop_config.json ; Cursor : .cursor/mcp.json ; VS Code et les autres utilisent la même forme « mcpServers »).
rail-mcpclaude mcp add rail -e 'RAIL_DATA_DIR=<RAIL_DATA_DIR>' -- npx -y rail-mcp@0.1.1
{
"mcpServers": {
"rail": {
"command": "npx",
"args": [
"-y",
"rail-mcp@0.1.1"
],
"env": {
"RAIL_DATA_DIR": "<RAIL_DATA_DIR>"
}
}
}
}
Variables d'environnement
RAIL_DATA_DIRDirectory for the local ledger (budget, proposals, receipts). Defaults to ~/.rail in the user home directory. Set to override.
Documentation
Tirée du README du projet sur GitHub, actualisée chaque semaine. Les liens ouvrent GitHub.
Sommaire
Rail MCP
Rail is a budget, approval and receipt layer that agents call when they buy something on a human's behalf, dry-run by default, wrapping Stripe Link for settlement rather than issuing cards.
It is not a marketplace, storefront, or outreach tool. Rail owns the budget, the approval gate, and the receipt ledger.
Quickstart
Requires Node.js 20+. Run the published package with npx -y rail-mcp. Default mode is dry-run: no network and no charges.
Claude Desktop and any other host that takes an mcpServers block:
{
"mcpServers": {
"rail": {
"command": "npx",
"args": ["-y", "rail-mcp"]
}
}
}
Cursor: one-click install (install links). The config value is the base64 of {"command":"npx","args":["-y","rail-mcp"]}.
Install Rail in Cursor
cursor://anysphere.cursor-deeplink/mcp/install?name=rail&config=eyJjb21tYW5kIjoibnB4IiwiYXJncyI6WyIteSIsInJhaWwtbWNwIl19
Grok Bot: add a custom MCP with npx -y rail-mcp.
Do not put Stripe or Link secrets (STRIPE_SECRET_KEY, LINK_ACCESS_TOKEN, or live-spend flags) in this shared snippet. Real charges stay off unless those gates are set on purpose, outside the shared config. See Stripe Link seam.
Ledger files (budget.json, proposals.json, receipts.json) are written to ~/.rail in the user home directory, or to RAIL_DATA_DIR when that is set. They are local state, not part of the npm package. The default does not follow the process working directory, so a host that starts the server from / still writes under the home directory.
Example
Set a budget, propose a purchase, have a human approve it, then read the receipt. Dry-run moves no money off the machine.
set_budget→{ "amount_usd": 50, "note": "week1" }propose_purchase→{ "merchant": "Acme", "amount_usd": 12, "rationale": "need widgets" }get_budgetstill shows remaining50.00andopen_proposals: 1. The proposal is pending until a human decides.decide_proposal→{ "proposal_id": "prop_…", "decision": "approve" }get_receiptsreturns onesettled_dry_runreceipt with asettlement_ref. Remaining budget is38.00.
Rejecting spends nothing. refund_receipt reverses a dry-run settlement and restores the budget.
Tools
| Tool | Purpose |
|---|---|
set_budget | Overwrite the local budget limit. Moves no money |
get_budget | Read budget, spent, remaining, open proposals count |
propose_purchase | Record a pending intent only. Moves no money. Needs a later human approval |
list_proposals | Read proposals by pending (default) / approved / rejected / all |
decide_proposal | Human confirmation step. approve settles (dry-run receipt + settlement_ref, remaining decremented). reject spends nothing. No double-decide |
get_receipts | Read receipts, newest first |
refund_receipt | Human confirmation step. Reverse a dry-run settlement and restore budget |
Permission model
Proposing a purchase is cheap. Settling one is not. An agent may record intent without moving money. The human should confirm the call that settles a purchase or reverses a receipt.
Hosts should read title, description, and annotations from tools/list. Every hint is set explicitly. The spec defaults (readOnlyHint false, destructiveHint true, openWorldHint true) would otherwise treat every tool as a destructive open-world write. Rail's ledger is local. Default mode is dry-run: no network and no charge.
| Tool | readOnlyHint | destructiveHint | idempotentHint | openWorldHint | Host should |
|---|---|---|---|---|---|
set_budget | false | false | true | false | Allow. Overwrites the local limit only. The same arguments do not stack or spend |
get_budget | true | false | true | false | Allow. Read-only |
propose_purchase | false | false | false | false | Allow. Adds one pending intent and spends nothing. Each call creates a new proposal |
list_proposals | true | false | true | false | Allow. Read-only |
decide_proposal | false | true | true | false | Confirm with the human. approve settles and decrements remaining budget. A repeat does not settle again |
get_receipts | true | false | true | false | Allow. Read-only |
refund_receipt | false | true | true | false | Confirm with the human. Reverses a dry-run settlement and restores budget. A repeat does not refund again |
destructiveHint is the confirmation signal, and it is true only for decide_proposal and refund_receipt. openWorldHint is false on every tool: dry-run never leaves the ledger, and live Link stays behind separate environment gates. Annotations are hints for the host. They do not themselves move money.
Mode defaults to RAIL_MODE=dry_run. Money is stored as integer cents; tools display USD with 2 decimals. IDs: prop_…, rcpt_…, dry-run Link refs lsrq_dry_… on settlement_ref.
Local checkout
tsx is a dev dependency for dogfood in this repo. npm run smoke runs the TypeScript sources directly and does not need a build. npm start compiles first, then runs dist/index.js (the same file the rail-mcp bin points at).
npm install
npm start # tsc → dist/, then stdio MCP server
npm run dev # tsx watch src/index.ts
npm run smoke # dry-run ledger smoke test; throwaway dirs only (never ./data, ~/.rail, or RAIL_DATA_DIR)
npm run smoke forces RAIL_MODE=dry_run, never enables live charge gates, and writes the ledger only under temporary directories it creates and deletes. ./data, ~/.rail, and RAIL_DATA_DIR are left untouched even when RAIL_DATA_DIR is set in the environment. One check starts the server with its working directory at / and HOME pointed at a temp directory, then calls a tool. That call succeeds, and the ledger for that check is created under the temp home rather than /data or the real ~/.rail. An explicit RAIL_DATA_DIR still overrides the default in that launch.
Stripe Link seam
Rail asks Link for a one-time credential. Rail still decides budget and policy and writes the receipt. This seam does not issue cards and does not render UI.
RAIL_MODE=dry_run(default, including when unset):createSpendRequestreturns a fake Link spend-request id and statusdry_run_pending_human. No Stripe or Link HTTP. Approve stores that id assettlement_refon asettled_dry_runreceipt and decrements remaining budget.RAIL_MODE=livewithout gates: throwslive mode not enabled — set keys and get explicit human approvalunless bothSTRIPE_SECRET_KEYandRAIL_LIVE=1are set. The proposal stays pending and no receipt is written.- Live with those two gates: the
POST https://api.link.com/spend_requestsbody is scaffolded only (seesrc/stripeLink.ts). Nothing is sent unlessRAIL_ALLOW_LIVE_CHARGE=1. RAIL_ALLOW_LIVE_CHARGE=1: the only branch allowed to call Link, and only ifLINK_ACCESS_TOKENis also set.STRIPE_SECRET_KEYis a presence gate and is never sent. There is no Stripe Issuing card create. Docs: Link CLI, link-cli, Issuing for agents.- Live spend stays off until those env vars are set on purpose. A posted Link request is stored as
link_pending_humanand does not decrement the dry-run budget. CI andnpm run smokestay on dry-run and do not charge.
| Env | Role |
|---|---|
RAIL_MODE | dry_run (default) or live |
RAIL_DATA_DIR | Ledger directory. Default is ~/.rail in the user home directory. An explicit value overrides the default |
STRIPE_SECRET_KEY | Required for live. Not used to issue cards or as a Link bearer token |
RAIL_LIVE=1 | Explicit live approval alongside the secret key |
RAIL_ALLOW_LIVE_CHARGE=1 | Additional gate before any Link HTTP |
LINK_ACCESS_TOKEN | Link OAuth token. Required before the scaffolded POST actually runs |
Explicit non-goals
- No marketplace UI
- No outreach / messaging
- No card issuing
- No real charges in dry-run, CI, or smoke
License
MIT. Copyright 2026 Thomas Stockham.
Questions fréquentes
Qu'est-ce que Rail ?
Budgets, propose-then-approve, and receipts for agents that buy for a human. Dry-run by default.
Comment installer Rail ?
Il s'exécute sur votre machine. Dans Claude Code : claude mcp add rail -e 'RAIL_DATA_DIR=<RAIL_DATA_DIR>' -- npx -y rail-mcp@0.1.1. Pour les autres clients, collez la configuration JSON de cette page dans leurs réglages MCP.
Faut-il une clé API pour Rail ?
Son entrée au registre n'indique aucune clé ni secret obligatoire. Certaines fonctions peuvent tout de même demander un compte auprès du service auquel il se connecte.
Rail est-il open source ?
Oui. Son code est sur GitHub sous licence MIT.
Qui maintient Rail ?
Il est publié par tstockham96. AI.tn le référence sans le maintenir ni le recommander.
Issu du registre MCP officiel. Être listé n'est pas une recommandation : vérifiez le code et les permissions d'un serveur avant de lui donner accès à vos données.