Rhumb Mcp

io.github.supertrained由 GitHub 账号 supertrained 发布v0.8.2更新于 Oct 24, 2026

Agent gateway for Index discovery, AN Score evaluation, and Resolve governed execution

概览

AI 生成的概览

Rhumb 网关 MCP 服务器,让助手发现、评分、路由并执行第三方 API 能力,并提供回执、预算与消费统计。

可以做什么
  • 搜索并比较服务,查看 AN Score 明细、替代方案与已知故障模式
  • 浏览并解析能力,获得按评分排序的提供商及其健康状态、成本与认证方式
  • 预估并通过 Rhumb Resolve 执行能力调用,再获取签名执行回执
  • 检查凭据就绪状态,获取分步的凭据申请指引
  • 列出、查看并运行已发布的 Layer 3 配方(测试阶段,目录目前很少)
  • 管理消费上限、余额、充值链接、账单流水与用量遥测
适用场景
当助手需要为某项任务(邮件、抽取、生成、窄范围数据补充)挑选第三方 API,并通过一个统一路由、计量的网关调用它,而不想逐个接入各提供商时适用。发现与评分无需账号;执行、财务与运维工具需要受管密钥。
运行要求
通过 npx(Node.js)以 stdio 在本地运行。发现功能无需凭据。执行、财务与运维工具需要在 RHUMB_API_KEY 中提供受管 API 密钥;也可改用 x402 钱包地址与私钥按次付费。需要访问 Rhumb API 的网络连接。
安装前请注意
执行类工具会真实调用第三方服务并产生费用,因此提供预算与消费工具。x402 方式要求提供钱包私钥,将其写入配置风险很高。受管密钥与钱包凭据均属机密。可执行提供商数量有限,配方目录基本为空。

安装

在 SourceWeft 中

  1. 打开 控制台中的 Rhumb Mcp,将其添加到工作区。
  2. 为需要使用其工具的对话启用该服务。

Desktop only,通过 STDIO。 STDIO 服务会启动本地进程,因此需要 SourceWeft 桌面宿主。

其他 MCP 客户端

参照 仓库 中的启动说明。

README

Rhumb MCP Server

Agent-native tool intelligence for the Model Context Protocol.

Three execution layers: raw provider access (Layer 1), intelligent routing (Layer 2), and deterministic composed recipes (Layer 3, beta with a truthful public catalog). Every provider rated with the AN Score. Every execution produces a chain-hashed receipt.

Zero-config quickstart

No governed API key needed for discovery. Install and start immediately:

bash
npx -y --package rhumb-mcp@latest rhumb-mcp

Or add to Claude Desktop / Cursor / any MCP client:

json
{  "mcpServers": {    "rhumb": {      "command": "npx",      "args": ["-y", "--package", "rhumb-mcp@latest", "rhumb-mcp"]    }  }}

Ask your agent:

  • "Find me the best email API for agents"
  • "What's the AN Score for Stripe?"
  • "Check whether any Rhumb recipes are published yet"

Current callable scope

Rhumb is strongest today for research, extraction, generation, and narrow enrichment.

Treat it as capability infrastructure first, not as a general business-agent automation layer yet. Layer 2 is the real production surface today; Layer 3 is still beta with an intentionally sparse public catalog.

Discovery breadth is wider than current execution breadth: Rhumb scores 999 services and exposes 435 capability definitions, but current governed execution is concentrated in 28 callable providers.

Resolve mental model

  • Service = vendor Rhumb evaluates and compares
  • Capability = executable action your MCP client can route and execute
  • Recipe = deterministic multi-step workflow on top of capabilities (beta, sparse public catalog)
  • Layer 2 is the default path — discover → resolve → estimate → execute
  • Default auth for repeat traffic = governed API key or wallet-prefund on X-Rhumb-Key
  • Bring BYOK or Agent Vault only when provider control is the point
  • Use x402 only when zero-signup per-call payment is the point

Canonical onboarding map: https://rhumb.dev/docs#resolve-mental-model

What's new in v2.0.0

Rhumb Resolve — three execution layers:

LayerWhatHow
Layer 1Raw provider accessYou pick the provider. Escape hatch + trust anchor.
Layer 2Capability routingRhumb routes to the best-fit supported provider for the call, using AN Score, availability / circuit state, estimated cost, latency proxy, credential mode, and explicit policy constraints.
Layer 3Deterministic recipes (beta)Compiled DAG workflows. Multi-step, budget-enforced, content-firewalled. No published recipes yet.

New infrastructure:

  • Execution receipts (chain-hashed, HMAC-signed)
  • Route explanations (why this provider was chosen)
  • AN Score structural separation (read-only cache, auditable)
  • Billing event stream (chain-hashed, 15+ event types)
  • Trust dashboard inputs for observed availability, costs, and reliability
  • Recipe safety controls (content firewall, idempotency, nesting depth, fan-out rate limiting)
  • Kill switches (per-agent, per-provider, per-recipe, global with authenticated two-person auth)
  • Audit trail (append-only, chain-hash verification, export API)

Migration from 0.x

  • All v1 endpoints remain fully backward compatible — no breaking changes
  • New v2 endpoints available alongside v1
  • execute_capability now returns _rhumb_v2 metadata with attribution and receipts
  • New tools for recipe-catalog inspection, receipts, and telemetry
  • Update: npx -y --package rhumb-mcp@latest rhumb-mcp

Discovery tools (no auth, 6 tools)

ToolWhat it does
find_servicesSearch indexed Services by what you need them to do
get_scoreGet the full AN Score breakdown for a Service: execution quality, access readiness, autonomy level, tier label, and freshness
get_alternativesFind alternative Services, ranked by AN Score
get_failure_modesGet known failure patterns, impact severity, and workarounds for a service
discover_capabilitiesBrowse Capabilities by domain or search text
resolve_capabilityGiven a Capability ID, and optionally a credential mode, returns ranked providers with health status, cost per call, auth methods, endpoint patterns, execute guidance, and machine-readable recovery fields like recovery_hint.resolve_url, recovery_hint.credential_modes_url, and, when applicable, recovery_hint.alternate_execute_hint or recovery_hint.setup_handoff, plus typo recovery when the capability ID is wrong

Execution tools (auth required, 8 tools)

json
{  "mcpServers": {    "rhumb": {      "command": "npx",      "args": ["-y", "rhumb-mcp@latest"],      "env": {        "RHUMB_API_KEY": "rk_your_key_here"      }    }  }}

Get a governed API key at https://rhumb.dev/auth/login (GitHub, Google, or email — 30 seconds).

ToolWhat it does
execute_capabilityCall a Capability through Rhumb Resolve
estimate_capabilityEstimate the active execution rail, cost, and health before a Capability call; anonymous direct system-of-record paths also preserve machine-readable execute_readiness handoffs
credential_ceremonyGet step-by-step instructions to obtain API credentials for a Service
check_credentialsInspect live credential-mode readiness, globally or for a specific Capability
rhumb_list_recipesList the current published Rhumb Layer 3 recipe catalog
rhumb_get_recipeGet the full published definition for a Rhumb recipe, including input/output schemas and step topology
rhumb_recipe_executeExecute a published Rhumb Layer 3 recipe once one is live in the public catalog
get_receiptRetrieve an execution receipt by ID

Financial tools (auth required, 5 tools)

ToolWhat it does
budgetCheck or set your call spending limit
spendGet your spending breakdown for a billing period: total USD spent, call count, average cost per call, broken down by Capability and by provider
check_balanceCheck your current Rhumb credit balance in USD
get_payment_urlGet a checkout URL to add credits to your Rhumb balance
get_ledgerGet your billing history: charges (debits), top-ups (credits), and auto-reload events

Operations tools (auth required, 2 tools)

ToolWhat it does
routingGet or set how Rhumb auto-selects providers when you don't specify one in execute_capability
usage_telemetryGet your execution analytics — calls, latency, errors, costs, and provider health for your Rhumb usage

21 MCP tools

Discovery (free): find_services, get_score, get_alternatives, get_failure_modes, discover_capabilities, resolve_capability

Execution (auth): execute_capability, estimate_capability, credential_ceremony, check_credentials, rhumb_list_recipes, rhumb_get_recipe, rhumb_recipe_execute, get_receipt

Financial (auth): budget, spend, check_balance, get_payment_url, get_ledger

Operations (auth): routing, usage_telemetry

Discovery spans 999 scored services, but current governed execution spans 28 callable providers.

Best current fit: research, extraction, generation, and narrow enrichment. Treat general business-agent automation as future scope, not the current callable promise.

Common workflows

1) Discover tools (no auth)

"I need an email provider for agents."

  • find_services → search the landscape
  • get_score → inspect a specific provider
  • get_failure_modes → see where it breaks in practice

2) Route a capability (no auth)

"I need email.send. What should I use?"

  • discover_capabilities → find the capability ID
  • resolve_capability → get ranked providers, optional credential-mode filtering, machine-readable recovery fields like recovery_hint.resolve_url, recovery_hint.credential_modes_url, and, when applicable, recovery_hint.alternate_execute_hint or recovery_hint.setup_handoff, or search suggestions when the capability ID is wrong

3) Check readiness (auth required)

"Can I call deployment.list right now, and on which rail?"

  • check_credentials → call without params for account-wide configured BYOK/direct-bundle readiness, or pass a capability to inspect provider-level mode status and ceremony availability

4) Execute (auth required)

"Send the email with the cheapest provider above my quality floor."

  • estimate_capability → check the active execution rail, health, and cost before execution
  • execute_capability → perform the action
  • get_receipt → verify the HMAC-signed execution record

5) Check recipe availability / run a recipe (auth required)

"Is there already a published Rhumb workflow for this?"

  • rhumb_list_recipes → check what is actually live in the public catalog
  • rhumb_get_recipe → inspect a recipe only after it appears there
  • rhumb_recipe_execute → run the compiled recipe once it is published
  • Public note: the Layer 3 catalog is currently sparse/empty, so most real work should use Layer 2 capabilities today

x402 micropayments

No account needed — pay per call with USDC on Base:

json
{  "env": {    "RHUMB_X402_WALLET_ADDRESS": "0x...",    "RHUMB_X402_PRIVATE_KEY": "0x..."  }}

Local development

bash
cd packages/mcpnpm cinpm run dev

Test and build

bash
npm test          # 84+ testsnpm run type-checknpm run build

Related

License

MIT

来源:packages/mcp/README.md,提交 957abf9

工具

0
SourceWeft 尚未在沙箱中启动过该软件包,因此其工具未知。

版本历史

1
  1. v0.8.2最新Sep 16, 2026

评分

已满足适用检查项的 50%

该评分衡量此收录条目的完整程度,依据注册中心条目、端点与公开来源计算得出。它不是对服务器本身的评价,也不表示服务器安全或当前可用:可连通性是带日期的检查结果,评分也不包含安装量或用户评价。

  • 近 30 天内可连通25 分

    请保持端点能够响应 MCP 握手;SourceWeft 每 30 天探测一次。

  • 已声明认证方式10 分

    请在注册中心条目中声明凭据请求头,或保持端点能够响应探测,以便判定其认证方式。

  • 已列出工具15 分

    请让服务器在无凭据时响应 tools/list,或发布可在沙箱中无凭据启动的软件包。

  • 描述必需15 分
  • 图标10 分
  • 文档15 分
  • 许可证10 分