sh-mcp/packages/shared/src/mcp.ts

103 lines
3.7 KiB
TypeScript
Raw Normal View History

Phase 1: runnable MCP + OpenAPI servers (ops + finance) (#3) * Add shared transport: dispatch, MCP + OpenAPI adapters, local auth Add the single authoritative tool-execution path (executeTool) plus the two universal interfaces over it (design.md §2.5, §7.3): - dispatch.ts: scope enforcement, ajv input validation, rate limiting, finance egress redaction (redactDeep), and structured audit emission on one path. - audit.ts / rate-limit.ts: injected AuditLogger + RateLimiter abstractions. - mcp.ts: low-level MCP Server with scope-filtered tools/list (tool-hiding) and tools/call routed through executeTool. - http.ts: Express host mounting /mcp, /openapi.json, POST /tools/:name, /healthz. - openapi.ts: buildOpenApiDocument wraps the existing path generator into a full OpenAPI 3.1 document. - local-auth.ts: LocalAuthProvider (dev bearer tokens) that refuses to construct outside SH_MCP_ENV=local and enforces audience binding (design.md §3, §6). * Add in-memory dev clients; make package tool exports lazy Add an in-memory Client implementation per integration package (seeded fake data, no network) selected when SH_MCP_ENV=local (build-plan §4). Gmail/calendar/ tasks dev clients partition by ctx.sub; payments/qbo seed sensitive-looking fields so the redaction egress path has real targets to mask. Make the eager default-tool exports in tasks/reminders/qbo LAZY (getDefaultTools) so importing a package barrel no longer constructs an AWS client at module load (build-plan §7 'no I/O at import time') — the previous eager construction broke server startup. Fix payments tsconfig rootDir (src, was '.') so its declarations resolve under dist/index.d.ts like the other 8 packages. * Add runnable sh-mcp-ops and sh-mcp-finance servers Two thin composition-root servers over the shared transport (design.md §3): - ops: internal-data, knowledge-base, google-maps, gmail, calendar, tasks, reminders. finance: qbo, payments (audited + redacted on egress). - config from env only (no hardcoded ids/issuer/tables); SH_MCP_ENV selects LocalAuthProvider + dev clients (local) vs CognitoAuthProvider + real stubs (aws). Finance applies the 15-min finance-token TTL ceiling (design.md §2.5). - index.ts is the only place .listen() is called; a Lambda handler placeholder is exported but not depended on. - synth-only CDK stubs (no real IAM/Cognito/WAF) so 'cdk synth' has a valid app (build-plan §6); READMEs document local run, dev tokens, curl, MCP Inspector. * Add security-weighted test suite + coverage gate; wire tooling Add tests for the highest-risk surface (build-plan §5, design.md §7.3): tool-hiding, server-side scope enforcement (incl. forced hidden calls), audience binding, input-schema validation, finance redaction on egress, audit emission with hashed args, prompt-injection regression (tool output is data), rate limiting, MCP conformance (in-memory transport round-trip), OpenAPI 3.1 validity, and local-auth safety. Add HTTP integration tests (supertest) for both servers and per-package dev-client tests. 405 tests pass. Wire the coverage gate into vitest.config.ts: 80% overall, with per-file thresholds on the auth + dispatch crown jewels; exclude deferred real client stubs, entrypoints, cdk apps, and aws-only config from the gate (documented). Extend eslint flat config + add .prettierignore to cover servers/. Commit the updated package-lock.json. * Suppress pre-existing dev-tooling + out-of-scope scanner findings Add written-justification suppressions for the 4 confirmed crit/high pre-push scanner findings, none of which are in this PR's Phase 1 production code: - npmaudit vitest / @vitest/coverage-v8 / vite: dev/test-only deps that never run in the deployed server/Lambda runtime (pins carried from Phase 0b; Dependabot will bump). - gitleaks docs/agentforce-plan.md secret: that file is not on this branch and not in this changeset; flagged for the maintainer to scrub on its own branch. The deep agentic /sh-security-review (required for this auth/authz-touching PR) was NOT run by the agent and is flagged outstanding in the PR body. * Address CodeQL findings: bound ajv error work + edge rate limiting GHAS code-scanning alerts on this PR: - dispatch.ts (js/resource-exhaustion): ajv ran with allErrors:true on untrusted input, letting a crafted payload force unbounded error enumeration. Switch to allErrors:false (default) so validation short-circuits on the first failure; the 400 still names that path. - http.ts (js/missing-rate-limiting): the authenticated routes (/mcp, /tools/:name) had no edge throttle — auth/JWT verification ran on every request before the per-sub dispatch limiter could apply. Add an IP-keyed express-rate-limit in front of authenticate (120/60s default, configurable), returning the standard 429 shape. Defense-in-depth over the per-sub + per-tool limiter in executeTool; API GW/WAF remains the production edge. Tests: +2 cases proving the edge limiter throttles before auth (429, not 401) on /tools and /mcp. 407 pass; tsc/eslint/prettier clean. * Fix polynomial ReDoS in Bearer-token extraction (CodeQL js/polynomial-redos) extractBearerToken matched /^Bearer\s+(.+)$/ — \s and . both match a space, so the two quantifiers overlap and a crafted header can drive polynomial backtracking. Require the capture to start with a non-whitespace char (/^Bearer\s+(\S.*)$/), removing the ambiguity → linear match. Behavior is unchanged for real tokens; +2 regression tests. * Harden auth + finance redaction (sh-security-review confirmed mediums) Two confirmed medium findings from the agentic security review: - Fail-open SH_MCP_ENV: config defaulted to 'local' when the var was unset, so a deploy that forgot SH_MCP_ENV=aws would silently run LocalAuthProvider and accept static dev bearer tokens (dev-finance-admin -> finance:admin). Now fail-closed: SH_MCP_ENV must be explicitly 'local' or 'aws' or the server refuses to start. Plus an independent guard in LocalAuthProvider that refuses to construct in an AWS runtime (AWS_LAMBDA_FUNCTION_NAME / AWS_EXECUTION_ENV present), regardless of the env flag. - Finance egress redaction gap: redactDeep only wholesale-masked a sensitive key when its value was a scalar; an object/array under a sensitive key was recursed into, letting a bare nested value (e.g. {account:{number:...}}) escape the keyword-gated pattern matcher. Now the entire subtree under a sensitive key is masked. No current finance tool emitted such shapes (all flat strings), so this closes a latent hole in the universal safety net. +4 tests (subtree redaction, AWS-runtime guard). 411 pass; coverage gate green. Review also produced lows (memo free-text digits, unsalted argsHash, unauth /openapi.json by-design, session-cap no-reset by-design) tracked separately; 0 confirmed critical/high — review verdict PASS.
2026-06-26 13:33:21 -04:00
/**
* MCP (Model Context Protocol) transport over the shared registry.
*
* Builds a low-level `@modelcontextprotocol/sdk` `Server` per authenticated
* request, bound to the caller's `AuthContext`, exposing exactly two handlers:
*
* - `tools/list` returns ONLY the tools whose `requiredScope` is in the
* caller's scopes — server-side TOOL-HIDING (design.md §2.5). A finance-less
* caller never sees finance tools.
* - `tools/call` routes through {@link executeTool}, so scope enforcement,
* input validation, rate limiting, finance redaction, and audit are
* identical to the OpenAPI path. Hiding is convenience; the 403 from
* `executeTool` is the real boundary (a forced call to a hidden tool still
* fails).
*
* The low-level `Server` (not the Zod-based `McpServer`) is used deliberately:
* our tools carry JSON-Schema input schemas and we need per-caller dynamic tool
* lists, which the high-level helper does not support cleanly.
*/
import { Server } from '@modelcontextprotocol/sdk/server/index.js';
import {
CallToolRequestSchema,
ListToolsRequestSchema,
ErrorCode,
McpError,
} from '@modelcontextprotocol/sdk/types.js';
import { executeTool, InputValidationError, UnknownToolError } from './dispatch.js';
import { ScopeError } from './auth.js';
import { RateLimitError } from './rate-limit.js';
import { visibleTools } from './visibility.js';
import type { DispatchDeps } from './dispatch.js';
import type { ToolRegistry } from './registry.js';
import type { AuthContext } from './types.js';
/** Server identity advertised in the MCP handshake. */
export interface McpServerInfo {
name: string;
version: string;
}
/**
* Create a low-level MCP `Server` scoped to one authenticated caller.
*
* @param registry the server's tool registry
* @param ctx the authenticated caller (drives tool-hiding)
* @param deps audit + rate-limit dependencies for the dispatch path
* @param info MCP server identity for the handshake
*/
export function createMcpServer(
registry: ToolRegistry,
ctx: AuthContext,
deps: DispatchDeps,
info: McpServerInfo,
): Server {
const server = new Server(info, { capabilities: { tools: {} } });
// tools/list — scope-filtered (tool-hiding).
server.setRequestHandler(ListToolsRequestSchema, async () => ({
tools: visibleTools(registry, ctx).map((tool) => ({
name: tool.name,
description: tool.description,
inputSchema: tool.inputSchema as { type: 'object' },
})),
}));
// tools/call — single authoritative dispatch path.
server.setRequestHandler(CallToolRequestSchema, async (request) => {
const { name, arguments: args } = request.params;
try {
const output = await executeTool(registry, ctx, name, args ?? {}, deps);
return {
content: [{ type: 'text', text: JSON.stringify(output) }],
structuredContent: output as Record<string, unknown>,
};
} catch (err) {
throw toMcpError(err);
}
});
return server;
}
/** Map dispatch errors to MCP protocol errors (no secrets / stack traces). */
function toMcpError(err: unknown): McpError {
if (err instanceof UnknownToolError) {
return new McpError(ErrorCode.MethodNotFound, err.message);
}
if (err instanceof ScopeError) {
return new McpError(ErrorCode.InvalidRequest, err.message);
}
if (err instanceof InputValidationError) {
return new McpError(ErrorCode.InvalidParams, err.message);
}
if (err instanceof RateLimitError) {
return new McpError(ErrorCode.InvalidRequest, err.message);
}
if (err instanceof McpError) return err;
// Generic handler failure: do not leak the underlying message verbatim.
return new McpError(ErrorCode.InternalError, 'Tool execution failed.');
}