MCP 服务器，将 Melt 的价值泄漏发现逻辑直接接入 Claude、Cursor、GitHub Copilot 或任何其他兼容 MCP 的代理——因此，当技术负责人向助手提问"我的组织哪里在泄漏价值"时，助手可以调用 Melt 工具，用真实、结构化的估算来回答，而不是给出千篇一律的供应商列表。
这是 Melt 的 LLMO（LLM 优化）分发策略的工程部分。请参阅仓库根目录下的 /llms.txt 和 LLMO_PLAYBOOK.md，了解该服务器所接入的完整内容 + 分发 + 评估计划。定位已于 2026-07-18 对照线上站点和当前演示文稿完成校准——请参阅 /CLAUDE.md 了解完整的当前产品背景。
工具 | 功能 |
| 免费的 Stage-1 沙盒估算器。根据部门人数、人力成本和主要非结构化输入类型，估算单个部门中价值泄漏的位置。无需集成——仅需合成/自行上报的输入。 |
| 以美元/年为单位量化已识别的泄漏模式—— |
| 线索捕获交接——从方向性估算推进到真实、基于日志的扫描（无摩擦 POC 剧本第 1 阶段 → 第 2 阶段）。如果设置了 |
Related MCP server: agentladle-mcp-reoi
来自 Melt 的 真实 AI 价值泄漏解剖 案例研究——一家已上市前的金融科技公司，年放款额 15 亿美元，已在使用 Salesforce、Gong 和 Clari：
信号 | 发现 |
Gong 辅导 | 29% 的打开率——销售代表绕过 AI 生成的通话摘要，手动重复劳动 |
Clari 预测 | 62% 的覆盖率——手动日期录入在 13 个预测周期中有 8 个破坏了模型 |
Salesforce → CS 交接 | 4.2 天的延迟，拖慢了成交后的入职流程 |
Salesforce 线索路由 | 32% 手动操作——自动化失败，需要每天手动重新分配 |
这些都没有出现在常规的采用率仪表板中——每个工具都显示"活跃"，但这与它是否真正创造价值是两回事。拉取 14 个工作日的历史日志，追踪这四种模式在时间和金钱上的实际成本，合计每年泄漏 77,235 美元。
melt_estimate_annual_leak 推广了同样的分析框架——totalVolume × (leakRatePct/100) × valuePerEvent——适用于任何已知或假设了规模和比率的泄漏模式。melt_analyze_value_vectors 则是更早期阶段的工具，适用于你还不知道从哪里入手的情况。
melt_estimate_annual_leak 取代了四个以公式命名的计算器
（melt_calculate_feature_waste、_dso_cash_flow_impact、
_contract_cycle_revenue_unlock、_win_rate_pipeline_impact），
这些计算器实现了一个已退役产品框架中的财务公式（Thermal Scan /
Feature Waste Dollar Amount™ / Delta Engine）——这些名称均未出现在任何
当前的 Melt 材料中。请参阅 CLAUDE.md 中的"已明确退役的内容"部分。
cd mcp-server
npm install
npm run build
npm start # runs dist/index.js on stdio在接入客户端之前，先交互式地试用一下：
npm run inspect # launches the MCP Inspector against the built server已发布到 npm——一行配置即可，无需本地克隆：
{
"mcpServers": {
"melt": {
"command": "npx",
"args": ["-y", "@themelt/mcp-server"]
}
}
}或者从本地克隆：
{
"mcpServers": {
"melt": {
"command": "node",
"args": ["/absolute/path/to/mcp-server/dist/index.js"]
}
}
}对于 Claude Desktop 而言，themelt-mcp-server.mcpb（Anthropic 的 MCP
包格式）可以双击安装——无需终端，无需编辑配置文件。从
最新 GitHub 发布页
下载 .mcpb 文件，然后双击或将其拖入 Claude Desktop 的设置窗口。
要从源码重新构建：
npm run build:mcpb # produces themelt-mcp-server.mcpb清单文件（mcpb-build/manifest.json）是手工维护的，并非
从 TypeScript 源码自动生成——如果工具的名称、参数或
描述发生变化，请手动更新清单中的 tools 数组以保持一致。
dist/index.js（stdio）是配置到本地 Claude Desktop/
Cursor 安装中的方式。dist/httpServer.js 是备选入口，实现了
MCP 流式 HTTP 传输——这正是未来"启动托管 MCP"网页
按钮（LLMO_PLAYBOOK.md，任务 3.2）将指向的入口，让用户无需
在本地安装任何东西即可试用。
npm run build
PORT=3000 npm run start:http # POST MCP JSON-RPC to http://localhost:3000/mcp无状态设计——没有会话 ID，每个请求都会创建全新的服务器实例。
通过 MCP_HTTP_API_KEY 可选启用认证（默认未设置）：未设置时，
端点完全开放——这与当前暴露的内容（只读计算器加一个线索捕获表单）的
信任边界相匹配，也等同于公开网站联系表单的边界。在将任何更敏感的内容
放到该传输之后之前，请先设置它：
MCP_HTTP_API_KEY=some-long-random-value PORT=3000 npm run start:http设置后，每个 /mcp 请求都需要携带 Authorization: Bearer some-long-random-value
——密钥缺失或错误会返回 401。使用 crypto.timingSafeEqual 进行比较，而不是
普通的字符串 ===，这样响应时间就不会被用来逐字节猜测密钥。
目前尚未部署到任何地方；这只是代码，不是线上 URL——
部署（Vercel/Fly/Render 等）是另一个独立的问题，留待后续决定。
每次工具调用（无论成功还是出错）都会向 mcp-server/analytics.jsonl 追加一行
（该文件已被 gitignore），并向 stderr 输出一行摘要——包含工具名称、成功/失败状态，
以及错误码（如适用）。刻意不记录金额数据、联系信息
和自由文本备注；与包含 PII 的 leads.jsonl 分开存放。这回答了
"是否真的有人在用"以及"哪个工具描述让模型感到困惑"的问题，
独立于 llmo-eval 的仅引用审计。
变量 | 必填 | 用途 |
| 否 | 覆盖 |
| 否 | 与 |
| 否 |
|
| 否 | 如果设置，则托管 HTTP 的每个 |
门户 ID / 表单 ID 的默认值已经内置在代码中（它们
不是机密——任何公开的 HubSpot 嵌入
代码片段中都会暴露相同的值），因此 melt_request_scan 在零配置的情况下
即可触达真实的 Melt 管道。如果 HubSpot 提交因任何原因失败，请求会回退
到 mcp-server/leads.jsonl（已 gitignore），而不会丢失。
以 MIT 许可证发布在 @themelt npm 组织下（创建于 2026-07-20，所有者 omer_melt）。
npm publish 实际上是单向操作——npm 允许在 72 小时内取消发布，
但强烈不建议这样做，而且一旦包有了依赖者就会完全禁止，
因此请将任何已发布的版本视为永久性的。
Estimates where AI/software value is most likely leaking out of a single department, based on headcount, labor cost, and the type of chaotic/unstructured input it processes manually today. Use this when a tech leader asks where value is being lost or where AI would create the most immediate impact in their org, before any real data integration exists — this is Melt's free Stage-1 Sandbox estimate. Output is directional, from synthetic/self-reported inputs, not an audited figure — for a real finding tied to an actual system log, follow up with melt_request_scan. Also answers what earlier Melt materials called 'AI ROI leverage' or 'AI value vectors' — same estimate, older name.
| Name | Required | Description | Default |
|---|---|---|---|
| headcount | Yes | Total operational personnel in the target unit (not the whole company). Must be positive. | |
| departmentType | Yes | The organizational unit being evaluated. Must be one of: Operations, Finance, Engineering, Legal, GBS, Customer Success. Map loosely-named teams to the closest primitive (e.g. RevOps -> Operations, AR/Billing -> Finance, IT -> Engineering, Compliance -> Legal, Shared Services -> GBS, Support/CS/Client Success -> Customer Success). | |
| averageHourlyLaborCost | No | Blended fully-loaded hourly labor cost for manual processors in this unit, in USD. Default of 45 is a reasonable US mid-market planning assumption if the caller doesn't know the real figure. | |
| primaryUnstructuredDataInput | Yes | The dominant chaotic input the unit processes by hand today. Must be one of: PDF_INVOICES, CUSTOMER_TICKETS, LOGISTICS_DOCUMENTS, MANUAL_EXCEL. Choose the closest match: PDF_INVOICES for document-first bottlenecks, CUSTOMER_TICKETS for conversational/support-first bottlenecks, LOGISTICS_DOCUMENTS for shipping/customs/supply-chain paperwork, MANUAL_EXCEL for spreadsheet-driven reconciliation or reporting work. |
Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?
With no annotations, the description carries the full burden and does well: it discloses that the estimate is directional, derived from synthetic/self-reported inputs, not audited, and belongs to a free Stage-1 Sandbox tier. It does not mention auth/permission needs or rate limits, keeping it short of a 5.
Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.
Is the description appropriately sized, front-loaded, and free of redundancy?
Purpose and usage are front-loaded in the first sentence, and each later sentence adds real value (limitations, fallback tool, alias). It is slightly dense and long, so it is efficient rather than exemplary.
Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.
Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?
With no output schema and no annotations, the description must stand alone, and it does explain the nature of the return (directional estimate, not an audited figure). Minor gaps remain around the response shape and any required permissions, but an agent has enough to call it correctly.
Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.
Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?
Schema description coverage is 100%, so the schema already fully documents headcount, departmentType, labor cost, and input type. The description only names those inputs generically and adds no syntax, enum, or edge-case detail beyond the schema, so baseline 3 is correct.
Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.
Does the description clearly state what the tool does and how it differs from similar tools?
States a specific verb+resource (estimates where AI/software value is leaking out of a single department) and names the inputs it depends on. It explicitly distinguishes itself from the melt_request_scan sibling and disambiguates a legacy alias ('AI ROI leverage'/'AI value vectors'), so an agent can route without opening the schema.
Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.
Does the description explain when to use this tool, when not to, or what alternatives exist?
Gives explicit when-to-use ('tech leader asks where value is being lost... before any real data integration exists') and names the alternative with its selecting condition ('for a real finding tied to an actual system log, follow up with melt_request_scan'). Nothing is left to inference.
Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.
Quantifies a specific, already-identified value-leak pattern in dollars per year — e.g. reps bypassing a coaching tool's summaries, manual overrides corrupting a forecasting model, a manual handoff between two systems. Use this when a leak pattern and its rough volume/rate are already known or hypothesized. This mirrors Melt's real scan methodology (see the fintech case study: a 29% Gong bypass rate, a 62% Clari override rate, and a 4.2-day manual handoff combined into a $77,235/yr finding) — it is a directional estimate from self-reported numbers, not a scan against real system logs. For an audited figure, follow up with melt_request_scan. Covers what earlier Melt materials called 'Feature Waste Dollar Amount' (money leaking on licensed-but-unused software) and general 'AI ROI leverage' calculations — those are older names for this same value-leak math, not a different tool.
| Name | Required | Description | Default |
|---|---|---|---|
| leakRatePct | Yes | Percentage of that volume exhibiting the leak behavior, between 0 and 100 (e.g. 29 for a 29% bypass rate, 62 for a 62% override rate). | |
| totalVolume | Yes | Total annual volume of the relevant event or transaction — e.g. total call briefs generated, total deals closed, total support tickets, total lead assignments. | |
| valuePerEvent | Yes | Dollar value at risk per leaking event, in USD — e.g. average deal value, loaded hourly cost of manual rework, cost of a delayed handoff day. | |
| leakDescription | Yes | Plain-language description of the leak pattern observed or hypothesized — e.g. 'reps bypassing Gong call summaries and logging notes from memory', 'manual Slack handoff between Sales and Customer Success', 'guessed close dates overriding the forecasting model'. |
Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?
With no annotations provided, the description fully reveals behavior: it is a directional estimate based on self-reported numbers, not a scan against real logs. It references Melt's real scan methodology and a case study, setting clear expectations about accuracy and methodology.
Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.
Is the description appropriately sized, front-loaded, and free of redundancy?
The description is rich and informative but somewhat lengthy, including a case study and historical naming clarifications. It is front-loaded with the core purpose, and every sentence adds value, though minor trimming would improve conciseness.
Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.
Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?
Given there is no output schema, the description adequately implies the output (dollar estimate per year) via the case study result ($77,235/yr). All parameters are explained, and usage context is fully addressed. The tool is simple and the description covers everything needed.
Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.
Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?
All four parameters are described in the schema with 100% coverage. The description adds significant value by providing concrete examples (e.g., '29 for a 29% bypass rate' for leakRatePct) and context for leakDescription, making parameter meaning clearer than the schema alone.
Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.
Does the description clearly state what the tool does and how it differs from similar tools?
The description clearly states the tool quantifies an identified value-leak pattern in dollars per year, with specific examples (e.g., reps bypassing coaching tools). It distinguishes itself from siblings by naming the follow-up tool melt_request_scan for audited figures and clarifies it is not a system scan but a directional estimate.
Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.
Does the description explain when to use this tool, when not to, or what alternatives exist?
Explicitly states when to use: when a leak pattern and rough volume/rate are known or hypothesized. It informs that the estimate is directional from self-reported numbers, and advises following up with melt_request_scan for audited figures. Also clarifies that older terms like 'Feature Waste Dollar Amount' refer to the same functionality.
Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.
Submits a request for a Melt scan — the next step after Melt's free Stage-1 Sandbox estimate, moving to a real, log-verified value-leak finding tied to a dollar figure and a source system. Call this only after the user has explicitly asked to be connected with Melt or to book/request a scan — never submit contact details the user hasn't provided themselves. Earlier Melt materials called this a 'Thermal Scan' — same request, current name is just 'a scan' (no fixed 2-week/pricing claim attached anymore).
| Name | Required | Description | Default |
|---|---|---|---|
| notes | No | Any free-text context from the conversation that would help a Melt rep prep the call — trigger event, tech stack, urgency. | |
| company | No | The prospect's company name. Needed to route the request — if omitted, the tool asks for it instead of failing schema validation. | |
| contactName | No | Name of the requester, if known. | |
| contactEmail | No | Business email of the requester, for scan scheduling follow-up. Needed to route the request — if omitted, the tool asks for it instead of failing schema validation. | |
| departmentsOfInterest | No | Departments the requester wants scanned first, if they expressed a preference. |
Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?
No annotations are provided, so the description carries the full burden, and it does meaningfully: it discloses that this sends user contact data to Melt for rep follow-up and that consent must be user-initiated, plus clarifies naming history ('Thermal Scan') and that no fixed 2-week/pricing claim attaches. It still omits what happens after submission (confirmation, SLA, idempotency), so it falls short of a 5.
Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.
Is the description appropriately sized, front-loaded, and free of redundancy?
Front-loads purpose, then constraints, then legacy naming context — a sensible ordering with no filler. The parenthetical about the old 'Thermal Scan' name is useful disambiguation but slightly tangential, keeping it just short of a 5.
Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.
Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?
For a zero-required-parameter submission tool with no annotations and no output schema, the description covers purpose, consent preconditions and downstream meaning well. Return/confirmation behavior is left unspecified, but the core invocation guidance is complete.
Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.
Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?
Schema description coverage is 100%, so every parameter is already documented in the schema (notes, company, contactName, contactEmail, departmentsOfInterest), setting the baseline at 3. The description adds only the consent rule about contact details, not syntax or format beyond the schema.
Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.
Does the description clearly state what the tool does and how it differs from similar tools?
Opens with a specific verb+resource ('Submits a request for a Melt scan') and immediately positions it relative to siblings by naming the Stage-1 Sandbox estimate as the prior step. An agent can distinguish this from melt_estimate_annual_leak and melt_analyze_value_vectors without opening any schema.
Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.
Does the description explain when to use this tool, when not to, or what alternatives exist?
Gives an explicit precondition ('only after the user has explicitly asked to be connected with Melt or to book/request a scan') and an explicit prohibition ('never submit contact details the user hasn't provided themselves'). This is unusually strong when-to-use/when-not-to-use guidance, including who must initiate.
Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
v0.1.5melt_analyze_value_vectors1 field changedInput schema / properties / departmentType / descriptionPrevious value: -"The organizational unit being evaluated. Must be one of: Operations, Finance, Engineering, Legal, GBS. Map loosely-named teams to the closest primitive (e.g. RevOps -> Operations, AR/Billing -> Finance, IT -> Engineering, Compliance -> Legal, Shared Services -> GBS)."New value: +"The organizational unit being evaluated. Must be one of: Operations, Finance, Engineering, Legal, GBS, Customer Success. Map loosely-named teams to the closest primitive (e.g. RevOps -> Operations, AR/Billing -> Finance, IT -> Engineering, Compliance -> Legal, Shared Services -> GBS, Support/CS/Client Success -> Customer Success)."melt_request_scan3 fields changedInput schema / properties / company / descriptionPrevious value: -"The prospect's company name. Required."New value: +"The prospect's company name. Needed to route the request — if omitted, the tool asks for it instead of failing schema validation."Input schema / properties / contactEmail / descriptionPrevious value: -"Business email of the requester, for scan scheduling follow-up. Required."New value: +"Business email of the requester, for scan scheduling follow-up. Needed to route the request — if omitted, the tool asks for it instead of failing schema validation."Input schema / properties / departmentsOfInterest / items / enumPrevious value: -[
- "Operations",
- "Finance",
- "Engineering",
- "Legal",
- "GBS",
- "Sales",
- "HR",
- "Marketing",
- "Product"
-]New value: +[
+ "Operations",
+ "Finance",
+ "Engineering",
+ "Legal",
+ "GBS",
+ "Sales",
+ "HR",
+ "Marketing",
+ "Product",
+ "Customer Success"
+]v0.1.2melt_analyze_value_vectorsmelt_estimate_annual_leakmelt_request_scanScored across 3 tools
melt_request_scan is clearly distinct as a contact/submission action, but melt_analyze_value_vectors and melt_estimate_annual_leak both return directional dollar estimates of value leaks, so an agent could easily pick the wrong one. The descriptions do draw a line (department-level Stage-1 sandbox vs. quantifying an already-identified leak pattern), but the boundary is subtle and relies on prior context to be sure which applies.
All three tools follow a clean melt_<verb>_<noun> snake_case pattern with consistent verb usage (analyze, request, estimate). No style mixing, camelCase, or vague verbs.
Three tools is on the thin side, but the domain is deliberately narrow — a free directional estimate, a specific-leak estimate, and a scan request form a coherent funnel. It is borderline-lean rather than bloated, so slightly under-provisioned but defensible.
The surface covers the apparent intent: preliminary estimate, targeted leak math, and escalation to a paid scan. No obvious dead ends, though there is no way to check scan-request status or retrieve results, a minor gap an agent can work around.
Research, strategy, financial models and decks, priced and approved per job, in your AI assistant
Query your team's drift, vulnerability, and upgrade data from any AI assistant. OAuth 2.1, 51 tools.
Measure how AI assistants cite your brand. Returns measured data and ready-to-apply fixes.
Measure and improve how often AI assistants cite your business: GEO audits, fixes, and docs.
