@bounded-systems/verbspec-mcp
Turns any verb Registry into a real MCP server, where each verb becomes an MCP tool with its own Zod input/output schemas (zero drift).
README
@bounded-systems/verbspec-mcp
Turn any @bounded-systems/verbspec
verb Registry into a real MCP server.
verbspec authors each check/generator once as a typed VerbSpec (Zod
input/output + a business-meaning summary) and projects it to CLI, Anthropic,
OpenAPI, and OpenRPC surfaces — and to MCP tool descriptors. What it
deliberately does not ship is a server: the MCP wire protocol, the
transport, and result-wrapping. This package is that missing layer, built on the
official @modelcontextprotocol/sdk.
Hand it a registry, get a server:
import { serveStdio } from "@bounded-systems/verbspec-mcp";
import { registry } from "./registry";
await serveStdio(registry, { name: "spd", version: "0.1.0" });
Every verb becomes an MCP tool whose inputSchema/outputSchema are the
verb's own Zod schemas (zero drift), and whose handler runs the verb. Object-
shaped outputs are returned as MCP structuredContent (and advertised as an
outputSchema); other outputs degrade to text. A verb that throws surfaces as an
isError tool result, not a transport crash.
Where this sits
One verb spec, projected everywhere — and every MCP server in the org is the same shape: verbspec verbs → this base → an optional topic layer.
verbspec author each verb once (Zod input/output + a summary)
└── verbspec-mcp the base — any verb Registry → MCP tools (this package)
└── <topic>-mcp optional, domain-specific layers on top, e.g.:
• verified static-site / website builds — adds read-only,
Sigstore-verified static responses + a resource catalog
(today: @bounded-systems/static-mcp)
verbspec-mcp is the only mandatory piece: it just runs the verbs (checks,
generators, mutations — whatever they do). Anything domain-specific is an
option stacked on top — still nothing but verbspec verbs plus this base plus
its own middleware. "Serve a verified website build" is one such topic, not the
core; use the base alone to expose any live verb registry as tools.
Install
Published to both registries so either ecosystem can consume it:
deno add jsr:@bounded-systems/verbspec-mcp # Deno / JSR-native
npm install @bounded-systems/verbspec-mcp # Node / npm (+ Bun)
@modelcontextprotocol/sdk is a normal dependency; @bounded-systems/verbspec
is a peer — the server shares your project's verbspec instance (and its
zod), so there's no dual-package split.
API
interface McpServerOptions {
name?: string; // handshake server name (default "verbspec-mcp")
version?: string; // handshake version (default "0.0.0")
instructions?: string; // handshake instructions (how to use the tools)
filter?: (verb: AnyVerbSpec) => boolean; // restrict which verbs are exposed
deps?: () => unknown; // inject a shared deps slice into every verb's run
mapResult?: (out, verb, args) => ToolResult | Promise<ToolResult>; // shape the tool result
}
// Build the configured SDK server WITHOUT a transport — attach your own (e.g. HTTP),
// or register extra surfaces (resources) on the returned server.
function buildMcpServer(registry: Registry, opts?: McpServerOptions): McpServer;
// buildMcpServer + connectStdio. The one-liner most servers want.
function serveStdio(registry: Registry, opts?: McpServerOptions): Promise<void>;
// Connect a server you built yourself to stdio, with stdout hygiene. For servers that
// also register resources and can't use the serveStdio shortcut.
function connectStdio(server: McpServer): Promise<void>;
For an HTTP server, take buildMcpServer(...) and .connect() it to a
Streamable HTTP transport from the SDK.
Topic layers: deps + mapResult
A domain-specific server stacks on the base through two seams instead of forking
it. deps injects a shared capability slice into every verb's run(input, deps)
(dependency injection — one client, or a test mock, for all verbs). mapResult
takes control of how a verb's output becomes the tool result (content,
structuredContent, _meta, isError). For example, the verified static-site
layer threads its verifying client via deps and renders verified bytes +
provenance _meta via mapResult — while the base still owns registration,
schema validation, and running the verb.
stdout hygiene
serveStdio redirects console.log/info/debug to stderr. The SDK writes
JSON-RPC frames to process.stdout, so a verb that logs progress during run()
would otherwise interleave and corrupt the stream (in Bun, console.log writes
to fd 1 natively — patching process.stdout alone would miss it). Verbs must not
write to process.stdout directly; use stderr for diagnostics.
License
MIT
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。