wasm-mcp
Read-only, deterministic MCP server for the WebAssembly core specification, enabling querying of instructions, types, sections, proposals, and spec text via tools like instruction_get, section_get, and spec_search.
README
wasm-mcp
Model Context Protocol server for the WebAssembly core specification. SHA-pinned, read-only, deterministic — safe to host as a public unauthenticated endpoint.
Not affiliated with, endorsed by, or sponsored by the W3C WebAssembly Community Group or Working Group.
What it gives you
spec_version— the pinned upstream commit and package version.instruction_get— opcode bytes, category, introducing version, stack type signature, validation + execution prose anchors / URLs, and trap conditions (traps+can_trap), by mnemonic (i32.add) or binary opcode (0x6a).instruction_list— enumerate, filterable by category (numeric, vector, reference, parametric, variable, table, memory, control, ref, i31, struct, array, extern), introducing version, or prefix.instruction_search— ranked free-text search across mnemonics, categories, and opcodes.type_get— value types (number / vector / reference) and type forms (functype,limits,memtype, …) with defining prose.section_get— one spec clause by id / anchor (structure, validation, execution, binary, text), with prose, cross-references, SpecTec formal-rule references, and the rendered URL.section_list— navigate the clause tree by area or anchor prefix.spec_search— full-text search across anchors, titles, and prose.proposal_list— WebAssembly proposals and their phases (from the pinnedWebAssembly/proposalsrepo), filterable by status, phase, champion, or affected spec.
section_get, section_list, and spec_search take a spec
argument covering all three specs in the WebAssembly/spec repo:
core (default), js-api (the JavaScript embedding API), and
web-api (Web-platform integration). The instruction and type tools
are core-only.
Contract
Every tool is:
- Read-only. No state mutation, no writes outside an optional local cache.
- Deterministic. Same input → same output, over the pinned spec
commit recorded in
vendor/PINNED.txt. - No execution. Never compiles, validates-by-running, instantiates, or runs any WebAssembly or arbitrary code. Validation and reduction rules are returned as data.
- No auth, no secrets, no PII. Usable anonymously.
- No network at request time. All spec data is fetched and indexed at build time and baked into the package.
Install (stdio, local)
npx wasm-mcp
Wire it into any MCP client by adding the server to its config (the launch command is the same everywhere; only the config file differs):
{
"mcpServers": {
"wasm": {
"type": "stdio",
"command": "npx",
"args": ["wasm-mcp"]
}
}
}
Hosted Worker
The Cloudflare Worker in worker/ exposes the same tool
surface as the stdio package over streamable HTTP at a single
unauthenticated endpoint, rate-limited per source IP (30 req / 60 s):
https://mcp.xyzzylabs.ai/wasm/mcp
GET /wasm/health reports status and the pinned SHAs; GET /wasm/privacy
states the anonymous, no-storage posture. All spec data is bundled
into the Worker, so it does pure in-memory lookups — no storage, no
network at request time.
Releases & data refresh
The pinned commits live in vendor/PINNED.txt
and are reported by spec_version. A scheduled GitHub Actions
workflow (refresh.yml) SHA-diffs
the upstream repos daily; when a pin moves it re-pins, bumps the patch
version, and tags a release, which publishes the npm package
(release.yml) and redeploys the
Worker (deploy-worker.yml).
Maintainers:
- npm publish uses Trusted Publishing
(OIDC) — no
NPM_TOKEN. Configure it once on npmjs.com (wasm-mcp → Settings → Trusted Publisher → GitHub Actions: orgxyzzylabs, repowasm-mcp, workflowrelease.yml). - Worker deploy needs
CLOUDFLARE_API_TOKEN+CLOUDFLARE_ACCOUNT_ID, stored as environment secrets on thecloudflareGitHub Environment (not repo-wide) with amain+v*deployment rule — see Securing the deploy credentials. - For the refresh workflow's tag push to trigger release + deploy,
add a
WORKFLOW_PATPAT (contents: write+workflows) — without it, refresh still re-pins and tags, but you run release / deploy manually. (Same secret name tc39-mcp uses, so one PAT — or anxyzzylabsorg secret — can serve both repos.)
License
MIT — see LICENSE.
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。