MCP: full detail per token, single-transaction tools, one user per token

Tokens get a level (migration 0034). Summary stays the default; Full detail is
the person's explicit choice on Connect your AI, with a confirmation, and is
written into CLAUDE.md as the one exception to summary-only cloud AI. Full
tokens get the full tools plus list_transactions, update_transactions and
mark_transfer_pair; IBANs and long account numbers are shortened in everything
they return (ids stay intact).

Isolation tested with two users and an accepted partner share: RLS alone lets
the partner read the owner's shared rows, but every tool also filters on the
token's user, so nothing leaked and no change landed (own ids, rule id, names).

Also fixes the full spending_summary crashing on rows without a category.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
MangaValkandClaude Opus 5.5 committed 2026-11-19 21:36:12 +01:00
1 parent dee883472a
commit 059a1da375
10 files changed
+175 -26

No files matched your search

+1
View File
@@ -5,6 +5,7 @@ Personal finance app. Next.js 15 (app router), Supabase (sign-in, Postgres, row-
## Rules for any change
- **Bank data stays private.** Every AI path goes through `src/lib/llm/gate.ts`. A cloud model gets the summary only: rounded amounts, names replaced by tokens. Full detail goes only to a local model. Never add another path.
- One exception, the person's own explicit choice: an MCP token made with "Full detail" on Connect your AI (`src/app/api/mcp`, migration 0034). Summary stays the default, IBANs and account numbers stay shortened, and a token only ever reaches its own user's rows (their session, plus a `user_id` filter in every tool, so a partner's shared accounts never show).
- **No AI API keys.** AI runs on the person's own signed-in app (Claude Code, Codex, Gemini CLI, opencode), a local model, or their OpenRouter sign-in.
- **Tools stay off for prompts from a website.** The helper (`public/broke-connect.cjs`) only runs AI apps with no tools at all. Don't add one that can read files.
- **Every table has row-level security.** Secrets like `plaid_items.access_token` have no policies, so only the server's secret key reads them.
+6 -1
View File
@@ -184,7 +184,12 @@ brokeless is an MCP server too, so an AI app like Claude Code can read your mone
3. Start a new Claude Code session and ask, for example: "Which of my brokeless rules catch too much?"
It is summary only, like broke ai: rounded amounts, months not days, names as placeholders (⟨M1⟩, ⟨A1⟩), never single transactions. Changes act on your real data, and Claude says what it will change first. Revoke a token on the same page. With AI switched off in Settings the endpoint refuses.
When you make a token you choose what it sees:
- **Summary** (default): like broke ai. Rounded amounts, months not days, names as placeholders (⟨M1⟩, ⟨A1⟩), never single transactions.
- **Full detail**: real names, exact amounts, dates and single transactions with their id, so Claude can fix rows one by one (`list_transactions`, `update_transactions`, `mark_transfer_pair`). IBANs and account numbers stay shortened (`NL24…5339`). Your real transactions then go to Anthropic while you use it, so brokeless asks you to confirm that.
A token only ever reaches its own user: every call runs in that user's session and every tool filters on that user too, so a partner's shared accounts don't show either. Changes act on your real data, and Claude says what it will change first. Revoke a token on the same page. With AI switched off in Settings the endpoint refuses.
## How AI stays private
+2 -2
View File
@@ -9,9 +9,9 @@ export async function POST(request: Request) {
const { data: { user } } = await (await createClient()).auth.getUser();
if (!user) return NextResponse.json({ error: "Not signed in" }, { status: 401 });
if (!sameOrigin(request)) return NextResponse.json({ error: "Forbidden" }, { status: 403 });
const { label } = (await request.json().catch(() => ({}))) as { label?: string };
const { label, level } = (await request.json().catch(() => ({}))) as { label?: string; level?: string };
try {
return NextResponse.json({ token: await createMcpToken(user.id, typeof label === "string" ? label : "Claude") });
return NextResponse.json({ token: await createMcpToken(user.id, typeof label === "string" ? label : "Claude", level === "full" ? "full" : "summary") });
} catch (e) {
console.error("[broke] mcp token:", e instanceof Error ? e.message : e);
return NextResponse.json({ error: "Couldn't make a token. If this is new, restart brokeless so the database updates." }, { status: 500 });
+18 -8
View File
@@ -7,20 +7,27 @@ import { aiAllowed, SUMMARY_NOTE } from "@/lib/llm/gate";
import { mcpUser } from "@/lib/llm/mcp-tokens";
import { makeTools } from "@/lib/llm/shared";
import { manageTools } from "@/lib/llm/tools/manage";
import { maskNumbers, rowTools } from "@/lib/llm/tools/rows";
import { createAdminClient } from "@/utils/supabase/admin";
// brokeless as an MCP server (Streamable HTTP, stateless) for AI apps such as Claude Code. A personal token from
// Settings (migration 0033) signs it in as that user; their own session runs every query, so RLS applies.
// Summary only, always: an outside AI is a cloud model (lib/llm/gate.ts). AI switched off in Settings: refused.
const READ_ONLY = new Set(["spending_summary", "search_transactions", "list_recurring", "list_accounts", "list_holdings", "get_profile", "list_categories", "list_rules"]);
const DESTRUCTIVE = new Set(["delete_rule", "merge_accounts", "recategorize"]);
// Summary by default: an outside AI is a cloud model (lib/llm/gate.ts). Full detail only for a token the person made
// with Full detail on Connect your AI (migration 0034): their explicit choice, per token. AI switched off: refused.
// One user per token: every call runs in that user's own session (RLS), and every tool filters on their id too.
const READ_ONLY = new Set(["spending_summary", "search_transactions", "list_recurring", "list_accounts", "list_holdings", "get_profile", "list_categories", "list_rules", "list_transactions"]);
const DESTRUCTIVE = new Set(["delete_rule", "merge_accounts", "recategorize", "update_transactions"]);
const CHANGES = "Tools that change things act on the real data: say what you will change and how many transactions it touches before you do it, and preview merges first.";
const INSTRUCTIONS = `You are connected to the user's own brokeless, a personal finance app. ${SUMMARY_NOTE}
Tools that change things (create_rule, delete_rule, recategorize, find_transfers, merge_accounts) act on the real data: say what you will change and how many transactions it touches before you do it, and preview merges first.`;
${CHANGES}`;
const FULL_INSTRUCTIONS = `You are connected to the user's own brokeless, a personal finance app, with full detail: real names, exact amounts, dates and single transactions with their id (IBANs and account numbers are shortened). Amounts: positive is money out, negative is money in. Transfers between the user's own accounts (category kind "transfer") count as neither spending nor income.
${CHANGES}`;
const deny = (status: number, message: string) => Response.json({ jsonrpc: "2.0", error: { code: -32001, message }, id: null }, { status });
async function handle(request: Request) {
const userId = await mcpUser(request.headers.get("authorization")?.replace(/^Bearer\s+/i, "") ?? null);
const who = await mcpUser(request.headers.get("authorization")?.replace(/^Bearer\s+/i, "") ?? null);
const userId = who?.userId, full = who?.level === "full";
if (!userId) return deny(401, "Missing or revoked token. Make one in brokeless: Connect your AI, Connect an AI app (MCP).");
const { data: { user } } = await createAdminClient().auth.admin.getUserById(userId);
if (!user) return deny(401, "That token's account is gone.");
@@ -28,8 +35,10 @@ async function handle(request: Request) {
const session = await agentClient(await mintAgentToken(user));
if (!session) return deny(503, "Couldn't open a session. Try again.");
const tools = [...makeTools(session.supabase, user.user_metadata?.profile, user.id, "summary"), ...manageTools(session.supabase, user.id)];
const server = new Server({ name: "brokeless", version: "1.0.0" }, { capabilities: { tools: {} }, instructions: INSTRUCTIONS });
const tools = full
? [...makeTools(session.supabase, user.user_metadata?.profile, user.id, "full"), ...manageTools(session.supabase, user.id, true), ...rowTools(session.supabase, user.id)]
: [...makeTools(session.supabase, user.user_metadata?.profile, user.id, "summary"), ...manageTools(session.supabase, user.id)];
const server = new Server({ name: "brokeless", version: "1.0.0" }, { capabilities: { tools: {} }, instructions: full ? FULL_INSTRUCTIONS : INSTRUCTIONS });
server.setRequestHandler(ListToolsRequestSchema, async () => ({
tools: tools.map((t) => ({
name: t.name, description: t.description, inputSchema: z.toJSONSchema(t.schema) as { type: "object" },
@@ -42,7 +51,8 @@ async function handle(request: Request) {
const parsed = t.schema.safeParse(req.params.arguments ?? {});
if (!parsed.success) return { content: [{ type: "text", text: `Invalid arguments: ${parsed.error.message}` }], isError: true };
try {
return { content: [{ type: "text", text: await t.run(parsed.data as Record<string, unknown>) }] };
const text = await t.run(parsed.data as Record<string, unknown>);
return { content: [{ type: "text", text: full ? maskNumbers(text) : text }] };
} catch (e) {
console.error("[broke] mcp tool:", t.name, e instanceof Error ? e.message : e);
return { content: [{ type: "text", text: "That didn't work. Try again." }], isError: true };
+21 -6
View File
@@ -5,7 +5,7 @@ import { createClient } from "@/utils/supabase/client";
// Connect your AI: personal tokens for brokeless's MCP endpoint (api/mcp), for AI apps such as Claude Code.
// A new token shows once, with the command that adds it to Claude Code. Revoking deletes it: the app loses access at once.
type Row = { id: string; label: string; created_at: string; last_used_at: string | null };
type Row = { id: string; label: string; created_at: string; last_used_at: string | null; level: "summary" | "full" };
const when = (d: string | null) => (d ? new Date(d).toLocaleDateString("en-US", { month: "short", day: "numeric", year: "numeric" }) : "never");
export function McpTokens() {
@@ -13,7 +13,9 @@ export function McpTokens() {
const [fresh, setFresh] = useState<string | null>(null);
const [err, setErr] = useState("");
const [busy, setBusy] = useState(false);
const load = () => createClient().from("mcp_tokens").select("id, label, created_at, last_used_at").order("created_at").then(({ data }) => setRows((data ?? []) as Row[]));
const [level, setLevel] = useState<"summary" | "full">("summary");
const [agree, setAgree] = useState(false);
const load = () => createClient().from("mcp_tokens").select("id, label, created_at, last_used_at, level").order("created_at").then(({ data }) => setRows((data ?? []) as Row[]));
useEffect(() => { load(); }, []);
const url = typeof window === "undefined" ? "" : `${window.location.origin}/api/mcp`;
const command = fresh ? `claude mcp add --scope user --transport http brokeless ${url} --header "Authorization: Bearer ${fresh}"` : "";
@@ -21,7 +23,7 @@ export function McpTokens() {
async function make() {
setBusy(true);
setErr("");
const r = await fetch("/api/mcp-tokens", { method: "POST", headers: { "content-type": "application/json" }, body: JSON.stringify({ label: "Claude" }) });
const r = await fetch("/api/mcp-tokens", { method: "POST", headers: { "content-type": "application/json" }, body: JSON.stringify({ label: level === "full" ? "Claude (full detail)" : "Claude", level }) });
const b = (await r.json().catch(() => ({}))) as { token?: string; error?: string };
setBusy(false);
if (!r.ok || !b.token) return setErr(b.error ?? "Couldn't make a token.");
@@ -45,7 +47,7 @@ export function McpTokens() {
<section className="card p-6 space-y-3">
<div className="space-y-1">
<h2 className="card-title">Connect an AI app (MCP)</h2>
<p className="text-sm text-muted">Let an AI app like Claude Code read your money summary and fix things for you: rules, categories, transfers, merging accounts. It gets the same summary as broke ai: rounded amounts and placeholder names, never single transactions.</p>
<p className="text-sm text-muted">Let an AI app like Claude Code look at your money and fix things for you: rules, categories, transfers, merging accounts. A token only ever reaches your own data.</p>
</div>
{fresh && (
<div className="space-y-2 rounded-xl border border-line p-3">
@@ -62,14 +64,27 @@ export function McpTokens() {
<ul className="text-sm">
{rows.map((r) => (
<li key={r.id} className="flex items-center justify-between gap-3 border-t border-line py-2">
<span>{r.label} <span className="text-muted">· made {when(r.created_at)} · last used {when(r.last_used_at)}</span></span>
<span>{r.label} <span className={r.level === "full" ? "text-bad" : "text-muted"}>· {r.level === "full" ? "full detail" : "summary"}</span> <span className="text-muted">· made {when(r.created_at)} · last used {when(r.last_used_at)}</span></span>
<button type="button" className="text-bad underline underline-offset-2" onClick={() => revoke(r.id)}>Revoke</button>
</li>
))}
</ul>
)}
{err && <p className="text-sm text-bad">{err}</p>}
<button type="button" className="btn" disabled={busy} onClick={make}>{busy ? "Making..." : "Make a token"}</button>
<fieldset className="space-y-2 text-sm">
<legend className="font-medium pb-1">What the AI app may see</legend>
<label className="flex gap-2 cursor-pointer"><input type="radio" name="mcp-level" checked={level === "summary"} onChange={() => setLevel("summary")} />
<span><span className="font-medium">Summary</span> <span className="text-muted">· like broke ai: rounded amounts, months, placeholder names (⟨M1⟩), never single transactions.</span></span></label>
<label className="flex gap-2 cursor-pointer"><input type="radio" name="mcp-level" checked={level === "full"} onChange={() => setLevel("full")} />
<span><span className="font-medium">Full detail</span> <span className="text-muted">· real names, exact amounts, dates and single transactions, so it can fix them one by one. IBANs and account numbers are shortened.</span></span></label>
{level === "full" && (
<label className="flex gap-2 cursor-pointer rounded-xl border border-line p-3">
<input type="checkbox" checked={agree} onChange={(e) => setAgree(e.target.checked)} />
<span>I understand the AI app sends my real transactions to its maker (for Claude: Anthropic) while I use it.</span>
</label>
)}
</fieldset>
<button type="button" className="btn" disabled={busy || (level === "full" && !agree)} onClick={make}>{busy ? "Making..." : level === "full" ? "Make a full-detail token" : "Make a token"}</button>
</section>
);
}
+7 -6
View File
@@ -6,19 +6,20 @@ import { createAdminClient } from "@/utils/supabase/admin";
// "bk_" in front so it's recognizable in a config file.
const hash = (t: string) => createHash("sha256").update(t).digest("hex");
export async function createMcpToken(userId: string, label = "Claude") {
export type McpLevel = "summary" | "full";
export async function createMcpToken(userId: string, label = "Claude", level: McpLevel = "summary") {
const token = `bk_${randomBytes(32).toString("base64url")}`;
const { error } = await createAdminClient().from("mcp_tokens").insert({ user_id: userId, token_hash: hash(token), label: label.slice(0, 60) || "Claude" });
const { error } = await createAdminClient().from("mcp_tokens").insert({ user_id: userId, token_hash: hash(token), label: label.slice(0, 60) || "Claude", level });
if (error) throw error;
return token;
}
/** The user a token belongs to, or null. Notes when it was last used (at most once a minute). */
export async function mcpUser(token: string | null): Promise<string | null> {
/** The user a token belongs to and what it may see, or null. Notes when it was last used (at most once a minute). */
export async function mcpUser(token: string | null): Promise<{ userId: string; level: McpLevel } | null> {
if (!token || !/^bk_[\w-]{20,}$/.test(token)) return null;
const db = createAdminClient();
const { data } = await db.from("mcp_tokens").select("id, user_id, last_used_at").eq("token_hash", hash(token)).maybeSingle();
const { data } = await db.from("mcp_tokens").select("id, user_id, last_used_at, level").eq("token_hash", hash(token)).maybeSingle();
if (!data) return null;
if (!data.last_used_at || Date.now() - Date.parse(data.last_used_at) > 60_000) await db.from("mcp_tokens").update({ last_used_at: new Date().toISOString() }).eq("id", data.id);
return data.user_id as string;
return { userId: data.user_id as string, level: data.level === "full" ? "full" : "summary" };
}
+2 -1
View File
@@ -128,7 +128,8 @@ export function fullTools(supabase: SupabaseClient, profile: Record<string, unkn
const byCat: Record<string, number> = {};
const byMerchant: Record<string, number> = {};
for (const t of spend) {
byCat[t.category!.label] = (byCat[t.category!.label] ?? 0) + t.amount;
const cat = t.category?.label ?? "Other"; // money out with no category counts as spend (isSpend), so it has none
byCat[cat] = (byCat[cat] ?? 0) + t.amount;
byMerchant[displayName(t)] = (byMerchant[displayName(t)] ?? 0) + t.amount;
}
const round = (o: Record<string, number>) => Object.entries(o).sort((a, b) => b[1] - a[1]).map(([k, v]) => [k, Math.round(v * 100) / 100]);
+3 -2
View File
@@ -18,7 +18,8 @@ import { createAdminClient } from "@/utils/supabase/admin";
const tool = <S extends z.ZodObject>(t: { name: string; description: string; schema: S; run: (input: z.infer<S>) => Promise<string> }) => t as unknown as Tool;
const ok = (x: unknown) => JSON.stringify(x);
export function manageTools(supabase: SupabaseClient, userId: string): Tool[] {
/** full: a token the person set to Full detail. Rule text then shows as written instead of with placeholders. */
export function manageTools(supabase: SupabaseClient, userId: string, full = false): Tool[] {
const names = () => namesFor(supabase, userId);
const categories = async () => {
const { data } = await supabase.from("categories").select("id, key, label, kind");
@@ -52,7 +53,7 @@ export function manageTools(supabase: SupabaseClient, userId: string): Tool[] {
const key = new Map(cats.map((c) => [c.id, c.key]));
const out = [];
for (const r of (data ?? []) as (Rule & { created_at: string })[])
out.push({ id: r.id, match: r.match ? mask(r.match, n) : null, direction: r.direction ?? "both", category: r.category_id ? key.get(r.category_id) ?? null : null,
out.push({ id: r.id, match: r.match ? (full ? r.match : mask(r.match, n)) : null, direction: r.direction ?? "both", category: r.category_id ? key.get(r.category_id) ?? null : null,
renames: !!r.rename, hides: !!r.ignore, matches_now: (await pastMatchesWith(supabase, userId, r)).length });
return ok(out);
},
+111
View File
@@ -0,0 +1,111 @@
import "server-only";
import type { SupabaseClient } from "@supabase/supabase-js";
import { z } from "zod";
import { displayName, TXN_SELECT, type Txn } from "@/lib/data";
import type { Tool } from "@/lib/llm/shared";
// Single-transaction tools for an MCP token the person set to "Full detail" (Connect your AI, migration 0034).
// Never for a summary token. Rows come with their id so an AI can fix them one by one; writes run as the user (RLS),
// only on the columns the app itself lets them change (category, name, hidden).
const tool = <S extends z.ZodObject>(t: { name: string; description: string; schema: S; run: (input: z.infer<S>) => Promise<string> }) => t as unknown as Tool;
/** IBANs and long account numbers keep only their first and last characters, in everything a full-detail token sees. */
// Ids (UUIDs) stay exactly as they are: a tool needs them back.
const UUID_RE = /[0-9a-f]{8}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{12}/gi;
const maskPart = (text: string) =>
text
.replace(/\b([A-Z]{2}\d{2})[A-Z0-9]{4,}(\d{4})\b/g, "$1…$2") // NL24INGB0008165339 -> NL24…5339
.replace(/\b\d{5,}(\d{4})\b/g, "…$1"); // 210026951876 -> …1876
export const maskNumbers = (text: string) => {
const ids = text.match(UUID_RE) ?? [];
return text.split(UUID_RE).map(maskPart).reduce((out, part, i) => out + part + (ids[i] ?? ""), "");
};
const UUID = z.string().uuid();
export function rowTools(supabase: SupabaseClient, userId: string): Tool[] {
const cats = async () => ((await supabase.from("categories").select("id, key, label, kind")).data ?? []) as { id: string; key: string; label: string; kind: string }[];
const catId = async (k: string) => { const t = k.trim().toLowerCase(); return (await cats()).find((c) => c.key.toLowerCase() === t || c.label.toLowerCase() === t)?.id ?? null; };
return [
tool({
name: "list_transactions",
description: "Single transactions with their id, newest first: date, amount (positive = money out, negative = money in), name, merchant, your own name for it, category key and kind, account, hidden, tags. Filters combine. IBANs and account numbers are shortened.",
schema: z.object({
query: z.string().optional().describe("Text in the name, merchant or your own name for it"),
from: z.string().regex(/^\d{4}-\d{2}-\d{2}$/).optional(),
to: z.string().regex(/^\d{4}-\d{2}-\d{2}$/).optional(),
account: z.string().optional().describe("Account name (see list_accounts)"),
category: z.string().optional().describe('Category key or label, or "none" for rows without one'),
direction: z.enum(["in", "out"]).optional(),
min_amount: z.number().optional().describe("At least this much (either direction)"),
limit: z.number().int().min(1).max(500).default(100),
offset: z.number().int().min(0).default(0),
}),
run: async ({ query, from, to, account, category, direction, min_amount, limit, offset }) => {
let q = supabase.from("transactions").select(TXN_SELECT).eq("user_id", userId);
if (query) { const s = query.replace(/[%,()]/g, " "); q = q.or(`name.ilike.%${s}%,merchant_name.ilike.%${s}%,custom_name.ilike.%${s}%`); }
if (from) q = q.gte("date", from);
if (to) q = q.lte("date", to);
if (direction === "out") q = q.gt("amount", 0);
if (direction === "in") q = q.lt("amount", 0);
if (account) {
const { data } = await supabase.from("accounts").select("id, name, custom_name").eq("user_id", userId);
const a = (data ?? []).find((x) => (x.custom_name || x.name).toLowerCase() === account.trim().toLowerCase());
if (!a) return JSON.stringify({ error: `No account "${account}".` });
q = q.eq("account_id", a.id);
}
if (category) {
if (category.trim().toLowerCase() === "none") q = q.is("category_id", null);
else { const id = await catId(category); if (!id) return JSON.stringify({ error: `No category "${category}". See list_categories.` }); q = q.eq("category_id", id); }
}
if (min_amount) q = q.or(`amount.gte.${Math.abs(min_amount)},amount.lte.${-Math.abs(min_amount)}`);
const { data, error } = await q.order("date", { ascending: false }).order("id").range(offset, offset + limit - 1);
if (error) return JSON.stringify({ error: error.message });
return JSON.stringify(((data ?? []) as unknown as Txn[]).map((t) => ({
id: t.id, date: t.date, amount: Number(t.amount), name: t.name, merchant: t.merchant_name, your_name: t.custom_name, shown_as: displayName(t),
category: t.category?.key ?? null, kind: t.category?.kind ?? null, account: t.account ? t.account.custom_name || t.account.name : null, hidden: t.ignored || undefined, tags: t.tags?.length ? t.tags : undefined,
})));
},
}),
tool({
name: "update_transactions",
description: "Change specific transactions by id (from list_transactions): category, your own name for them, or hidden (left out of every total). Returns how many changed.",
schema: z.object({
ids: z.array(UUID).min(1).max(500),
category: z.string().optional().describe('Category key or label, or "none" to clear it'),
name: z.string().max(120).nullable().optional().describe("Your own name for them; null clears it"),
hidden: z.boolean().optional(),
}),
run: async ({ ids, category, name, hidden }) => {
const patch: Record<string, unknown> = {};
if (category !== undefined) {
if (category.trim().toLowerCase() === "none") patch.category_id = null;
else { const id = await catId(category); if (!id) return JSON.stringify({ error: `No category "${category}".` }); patch.category_id = id; }
}
if (name !== undefined) patch.custom_name = name?.trim() || null;
if (hidden !== undefined) patch.ignored = hidden;
if (!Object.keys(patch).length) return JSON.stringify({ error: "Nothing to change." });
const { data, error } = await supabase.from("transactions").update(patch).eq("user_id", userId).in("id", ids).select("id");
return JSON.stringify(error ? { error: "Couldn't update them." } : { updated: data?.length ?? 0 });
},
}),
tool({
name: "mark_transfer_pair",
description: "Mark two transactions as one transfer between your own accounts: the money-out row becomes Transfers out, the money-in row Transfers in, so neither counts as spending or income. They must be the same amount, in different accounts.",
schema: z.object({ out_id: UUID, in_id: UUID }),
run: async ({ out_id, in_id }) => {
const { data } = await supabase.from("transactions").select("id, amount, account_id").eq("user_id", userId).in("id", [out_id, in_id]);
const o = data?.find((t) => t.id === out_id), i = data?.find((t) => t.id === in_id);
if (!o || !i) return JSON.stringify({ error: "Transaction not found." });
if (!(Number(o.amount) > 0 && Number(i.amount) < 0) || Math.abs(Number(o.amount) + Number(i.amount)) >= 0.005) return JSON.stringify({ error: "out_id must be money out and in_id the same amount coming in." });
if (o.account_id === i.account_id) return JSON.stringify({ error: "They're in the same account." });
const ids = Object.fromEntries((await cats()).filter((c) => c.key === "TRANSFER_OUT" || c.key === "TRANSFER_IN").map((c) => [c.key, c.id]));
await supabase.from("transactions").update({ category_id: ids.TRANSFER_OUT }).eq("id", out_id).eq("user_id", userId);
await supabase.from("transactions").update({ category_id: ids.TRANSFER_IN }).eq("id", in_id).eq("user_id", userId);
return JSON.stringify({ marked: true });
},
}),
];
}
@@ -0,0 +1,4 @@
-- What an MCP token may see: the summary (like every cloud AI), or full detail, which the person turns on for that
-- token themselves on Connect your AI. Tokens made before this stay summary.
alter table public.mcp_tokens add column if not exists level text not null default 'summary' check (level in ('summary', 'full'));
grant select (level) on public.mcp_tokens to authenticated;