
Supplier Invoice Duplicates
io.github.theluckystrike由 GitHub 账号 theluckystrike 发布v0.22.8更新于 Oct 24, 2026
Review duplicate supplier invoice candidates with grouped local evidence; no payment approval.
概览
本地 MCP 服务器,对用户提供的供应商发票批次做重复检查,按重复编号或相同金额与日期分组并返回证据。
- 可以做什么
- 对 1–1000 条提供的发票或贷项通知单记录执行重复检查,按完全相同的编号、规范化后的编号或相同金额与日期分组
- 列出、读取、分页浏览和删除已保存的报告,可按记录、分组、记录摘要和分组成员分节翻页
- 将报告导出为 CSV 文本,每条源记录一行,含原因代码和分组编号
- 查询离线授权等级并激活 Pro 许可证密钥
- 适用场景
- 当你已有发票历史导出数据,希望助手在付款前标出重复发票编号或金额与日期相同的记录以供人工复核时适用。它只处理每次提交的记录,不连接 ERP,也不读取 PDF。
- 运行要求
- 以本地 stdio 进程运行,需要 Node.js 18 或更高版本和 MCP 客户端。npm 包未发布,需从源码检出构建并在客户端填写绝对路径,或安装 Claude Desktop 的打包文件。免费使用无需 API 密钥或账号;MCP_LICENSE_KEY 为可选的 Pro 密钥。报告保存在本地数据目录。
安装
在 SourceWeft 中
- 打开 控制台中的 Supplier Invoice Duplicates,将其添加到工作区。
- 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
Duplicate supplier invoice check MCP server
Review a supplied batch of supplier invoices for repeated invoice numbers, conservative formatting variations, or identical amounts and invoice dates. This local MCP server returns complete grouped evidence and the exact reason each record needs review. Distinct legitimate invoices can share an amount and date. Candidates are not confirmed duplicates, fraud scores, recoverable money or payment decisions. It does not connect to an ERP, read PDFs, use OCR, or retrieve missing history.
Built by theluckystrike. MIT. Support: [email protected].
Try it before installing
Run the free duplicate-invoice example in your browser. Edit the synthetic invoice records to see candidate groups and their reasons. The example runs locally in your browser without an account or uploads. Product details and installation options.
Both servers include their full diagnostics and CSV evidence for free. Pro removes the limit of three saved reports; the $39 bundle covers both. Neither tool approves payment or retrieves missing accounting history.
Install
Claude Desktop: download duplicate-invoice-check.mcpb v0.22.8 and open it. SHA-256: 181379eb6dcff3d90b55e080dd9c180f6c96802e991872865a6b0ce422b3aaea. Other clients can use the source instructions below.
Requires Node.js 18 or newer and an MCP client. No API key or account is needed for free use. The npm package is not published: build the source checkout, then use an absolute local path.
Claude Desktop: merge this entry into mcpServers in ~/Library/Application Support/Claude/claude_desktop_config.json on macOS, or %APPDATA%\Claude\claude_desktop_config.json on Windows. Replace /absolute/path with the checkout location.
Claude Code:
Cursor: put this local entry in .cursor/mcp.json in your project or ~/.cursor/mcp.json:
VS Code: use .vscode/mcp.json with the top-level key servers:
Windsurf's legacy Cascade agent: merge the Claude Desktop mcpServers JSON into ~/.codeium/windsurf/mcp_config.json. This file applies to Cascade; its newer Devin Local agent uses a different configuration route. The standalone server exposes eight tools.
Cline: open its MCP Servers panel, Configure tab, then Configure MCP Servers. Add this entry to the settings JSON (for the Cline CLI, the file is ~/.cline/mcp.json):
These are the six local stdio client configurations used by this project. There is no hosted endpoint or browser chat connector. If the client cannot find node, replace it with the absolute path printed by which node. After saving, reconnect the server (fully quit and restart Claude Desktop), call license_status, then run the example below. A tools/list result alone does not verify candidate checking.
Tools
A repeated reference and a legitimate equal total
Call duplicate_invoice_check with synthetic records:
The result has four unique candidate records and three groups. exact_reference includes A/B despite their different dates and totals. normalized_reference includes A/B/C and exposes every conflicting field. same_amount_date includes A/D: two different legitimate services could have the same date and total, so this group is a review prompt. A appears in three groups but is counted once. No sum of possible losses is calculated. Changing C to INV-1 excludes it from the normalized INV001 group because leading zeros remain meaningful.
Exact input contract
Supply 1–1000 records, each with unique record_id, exact entity_id and supplier_id, document_type (invoice or credit_note), invoice_number, real YYYY-MM-DD invoice_date, uppercase three-letter currency, and positive safe-integer total_minor. IDs and invoice numbers are 1–160 characters without surrounding whitespace or control characters. Optional entity_label and supplier_label are display-only text, up to 200 characters without controls. Unknown fields are rejected.
Amounts are integer currency minor units, never floating point amounts: USD 12500 means $125.00. Currency codes are checked for uppercase three-letter syntax, not against an external currency registry. Credit notes use positive absolute amounts and compare only with credit notes. No amounts are added, converted or interpreted as a financial loss. Year 0000, impossible calendar dates, duplicate record ids, zero/negative/decimal/unsafe money and over-limit batches are input errors.
Each rule groups only records with exactly equal (entity_id, supplier_id, document_type, currency). Entity and supplier IDs are case-sensitive: similar names, different IDs and different entities never join. Multiple entities can appear in a batch, but each remains a separate scope. Display labels do not affect matching.
Three equality rules run independently:
exact_reference: the original invoice number matches exactly, even when amounts/dates differ.normalized_reference: ASCIIa-zbecomesA-Z; only ASCII space, hyphen, slash and dot are removed. Leading zeros, accents, all Unicode characters, underscores, nonbreaking spaces, Unicode dashes and letter/digit lookalikes remain unchanged. Empty normalized keys do not collide.same_amount_date: positive amount and invoice date both match exactly. This frequently has legitimate explanations.
Groups contain complete membership, exact scope and match key, an explanation and differing source fields. They are independent overlapping reasons, not transitive confirmed-duplicate clusters. candidate_record_count counts the union once. no_candidates_in_supplied_batch means these limited rules found none in the records supplied; it does not mean safe to pay. Missing history and reference, date or amount changes can evade these rules. Saved reports never become implicit cross-call history: include all records you want compared every time.
Source records and report groups are sorted deterministically; reordering the same input records produces the same diagnostic evidence. The computation groups keys instead of producing every pair: at most 3000 memberships for 1000 records, including overlap. The server performs no telemetry or network requests.
Free evidence, paging and CSV
Every check returns a retrieval.report_id, including unsaved checks. Preview IDs are preview:<UUID>, local to the current server process. The last two previews remain available for 30 minutes. A third check, restart or expiry can evict a preview; resubmit the original free diagnostic. Saved UUIDs persist on disk.
report_page sections are records, groups, record_summaries, and group_members. Follow next_offset until null, with limit 1–100. Group pages return summary metadata and member_count; read all members using section: "group_members" and the exact group_id. Member pages contain complete original records. Records are stored once in the report; groups refer to their exact record ids. All pages and CSV exports are free, including after hitting the saved-report cap.
The encoded incoming stdio message must fit the SDK's 10 MiB receive buffer. Responses use an 8 MiB escaped text-envelope budget; a larger report returns an explicit paginated overview with the same free retrieval instructions. Current 1000-row/text limits fit that budget in transport tests, including worst-case quoting. CSV uses one row per record rather than repeating it for each group. It includes original source fields, candidate flag, reason codes and group ids. Formula-like leading text is escaped. Export returns text; it never writes to a caller-supplied filename. CSV pages can contain fewer rows than the requested limit when quoting would exceed the response budget. Follow the actual next_offset until null; returned always counts the rows in that page. Every CSV page repeats its header; retain one header when joining pages.
Free and Pro
save: true at the free cap still returns the complete diagnostic and free preview with saved: false, save_error: "free_saved_report_cap", and an upgrade link. Delete a saved report to reuse a slot. Storage failures also preserve diagnostics. Pro removes only the saved-count cap; it does not increase input, transport, preview or total storage bounds.
Get Pro: $19 one time. All servers: $39 bundle. Set MCP_LICENSE_KEY or call license_activate. Keys verify offline.
All data stays local. Reports live under ${XDG_DATA_HOME:-~/.local/share}/mcp-servers/duplicate-invoice-check/. XDG_DATA_HOME must be absolute and directory ancestry must contain no symlinks. Private JSON files use atomic writes and cross-process mutation locks; reads revalidate source records and recompute the evidence instead of trusting edited derived results. License activation writes the shared local license config. Manage local accounting-data access and backups yourself.
Development
Build Docker from the repository root: docker build -f servers/duplicate-invoice-check/Dockerfile .. Tests cover actual stdio calls, schema rejection, scope/Unicode/zero negatives, an independent 2000-batch pairwise oracle, grouped 1000-record transport, pages and CSV, cross-process save caps, recomputation, symlink/path rejection and aggregate storage boundaries. A local signed-key stdio test uses the operator's existing signer if present; it skips when that private signer is absent.
来源:servers/duplicate-invoice-check/README.md,提交 0983f9c
工具
0版本历史
2- v0.22.8最新Oct 16, 2026
- v0.22.7Oct 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