claim-reconciliation-agent
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@claim-reconciliation-agent Split this platform payout by merchant and generate the 13-column claim template."
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
claim-reconciliation-agent
AI + MCP + Skill 的 ToB 内部效率产品:把「平台账单 → 企业 OA 回款拆分」这条纯人工财务认领链路,做成只读核验 + 可解释候选生成的自动化工具。
6 角色多 Agent 协作 | 文件单写者锁 | 全链路零写操作安全边界
它解决什么问题
平台按账期给商家打款,财务需要把「一笔到账流水」拆分到多个商家/酒店,再在 OA 里逐行认领。人工做这件事的痛点:流水号要从 OA 页面逐条抄、金额要按商家拆分对账、合同责任人要逐个查、抄错一个号就导入失败。
本项目的做法:本地 MCP Server(只读)+ 编排 Skill——用户人工登录 OA 并导出账单后,系统只读采集合同与银行流水、按商家精确拆分金额、生成可解释的认领候选,最后填入固定 13 列模板供人工复核导入。
产品边界(硬约束):不保存账号密码,不自动登录,不绕过验证码,不自动导入 OA,不执行财务核销。全程 write_actions=0。
Related MCP server: maxed-mcp
一真源 + 看板机制
collaboration 台账思想落地的「一真源」原则:同一份数据只允许一个权威出处,其余全部是脚本自动生成的派生物——手改派生物会被单测拦下。scripts/build_dashboard.py解析台账生成state.json+ 看板页;serve_dashboard.py本机起一个小站(3 秒轮询刷新);watch_dashboard.py纯标准库守护自动重建。看板含每个阶段的 owner、verifier、phase、next_action(演示图见
docs/preview.html,数据为脱敏重绘的演示数据)。
6 角色多 Agent 协作 + 单写者锁
AGENTS.md 治理 6 个角色:产品规划 / 项目经理 / 产品开发 / QA / Debug / 发布运营,每个角色绑定专属 Skill(grill-me、to-spec、implement、code-review、debug-agent 等)。
协作纪律踩过的坑,做成了机器可检查的机制:
文件单写者锁(
scripts/agent_file_lock.py):改共享文件前必须抢锁,抢不到必须 PAUSE,不许硬破别人的锁;30 分钟过期自动释放,PM 是唯一仲裁者。缘起是实测两个 Agent 并发改写同一文件导致提交覆盖。验收闭环:开发完成只能提交 REVIEW、不得自验;失败走 QA → Debug → QA 独立回归;失败/非唯一统一返回
PAUSED,永不出PASS。
零写边界与门禁
MCP 工具白名单只有 4 个只读能力(状态检查 / 就绪检测 / 流水号证据 / 模板生成到本地),禁登录、禁上传、禁导入、禁核销。
结构化阻塞协议
TASK_BLOCKED:失败/非唯一时返回run_id+blocker_fingerprint+ 证据摘要 + 允许动作,从最近安全阶段恢复。浏览器授权弹窗从「每步一次」压缩到「全程一次」(一次性授权 + 会话复用 + 闲置回收)。
生成侧门禁:酒店代码逐行稽核(错/空即非 0 退出)、模板缺号硬门禁、来源门禁(见下)。
反事故案例:8 个伪造流水号
一次交付中,交叉取证发现 8 个流水号是伪造的——与实际抽样交集为空、账期错位。处置:落地来源门禁脚本(scripts/verify_flow_provenance.py),把已交付物全部标记为不可用并撤回,之后每一笔流水号都必须带可回溯的来源证据。
这件事沉淀成一条原则:数据运营里,口径错误的代价远大于效率损失。
量化口径(均为项目真实记录):
单账期逐日对账 diff = 0.0
变异测试 18/18 全捕获(验证门禁真的有效,不是摆设)
交付治理 Issue DONE 24/31(≈77%)
累计 48 张已导入领款表(脱敏统计口径)
浏览器授权弹窗「每步一次 → 全程一次」
write_actions = 0 零写操作
快速开始
# 1) 安装依赖(Python 3.11+)
pip install -e .
pip install pytest
# 2) 跑测试(仓库内全部为合成假数据,无任何真实业务数据)
set PYTHONPATH=src # Windows
export PYTHONPATH=src # macOS / Linux
pytest
# 3) 从模板复制运行时配置(流水号留空即可体验解析路径)
cp scripts/_flow_numbers.json.template outputs/_flow_numbers.json
# 4) 重建看板(解析台账 → state.json + HTML)
python scripts/build_dashboard.py仓库内的 data/claiming_store.db、outputs/_flow_numbers.json、平台账单/*.xlsx 均为合成演示数据(纯假酒店/假商家/假金额),可安全运行。
仓库导航
路径 | 内容 |
| MCP Server 本体(合同读取、流水证据、模板生成、存储) |
| 编排入口 Skill(SKILL.md + 运行时契约) |
| 门禁、看板、单写者锁、模板生成等 26 个工具脚本 |
| 合成 fixture 支撑的完整测试套件 |
| 产品规格、技术设计、各环节 SOP |
| 6 角色协作协议与锁协议 |
免责与脱敏声明
本仓库是内部项目的脱敏发布副本:真实公司名、酒店名、商家名、人名、流水号、金额、账期一律未随仓发布;仓库内出现的示例数据全部为合成假数据。
原始台账、真实领款表与运行数据库仅保留在本地,永不推送。
已知债务如实登记:原内部仓的 xlutils 兼容问题在本副本中通过重建模板与合成 fixture 解决;依赖真实 OA 页面快照的回归测试以合成 fixture 等价覆盖。
本项目与任何平台/企业无隶属关系;模板格式仅用于演示「回款拆分」这一通用财务场景。
License
MIT
Available Tools
5 toolscheck_oa_readonly_readinessC
判断人工登录后的三项只读检查;全部通过才创建本次任务编号。
Inputs: bill_readable (required), left_login_page (required), connection_healthy (required), bank_flow_accessible (required). Returns: object
confirm_oa_flow_numbersC
生成列表查重校验号;不打开详情,发现冲突就暂停。
Inputs: rows (required), historical_keys. Returns: object
generate_claim_templateA
生成可导入 OA 的本地模板副本。
消费编排传入的结构化候选(每条含完整 32 位 hex 流水号、公司、账户、单位名称、责任人、认领状态),复用阶段一模板生成能力写出本地 XLS。完整流水号缺失或截断即结构化暂停;不连接/读取/写入/上传/导入真实 OA,不自行计算或回显完整流水号,不覆盖内置源模板。
Inputs: rows (required), output_path (required), source_template (required). Returns: object
open_greencloud_oa_commandB
返回让浏览器操作能力打开固定 OA 登录页的指令。
Inputs: no named parameters. Returns: object
startup_statusC
返回两个准备项,并说明是否可以开始只读检测。
Inputs: bill_readable (required), oa_window_opened (required). Returns: object
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 5 tools
The five tools target distinct workflow steps: opening OA login, startup readiness summary, post-login read-only checks, duplicate verification, and local template generation. Startup_status and check_oa_readonly_readiness both relate to readiness and could cause minor confusion about ordering, but descriptions clarify their different phases.
Four tools follow a clear verb_noun snake_case pattern (open_..., confirm_..., check_..., generate_...). startup_status is noun_noun and lacks a verb, which is a minor deviation but still readable and consistent in casing.
Five tools is well-scoped for a narrow claim-preparation workflow. Each tool covers a distinct preparation or validation step without redundancy, and the set is neither bloated nor thin.
The tools cover OA login, readiness checks, duplicate validation, and template generation, but there is no tool to perform actual claim matching, read OA claim data, or update/submit reconciliation results. These are notable gaps for a claim-reconciliation domain, though the agent explicitly avoids real OA writes.
Related MCP Connectors
Read-only MCP over your Excel or Airtable plus the PDFs beside it, joined as one relational dataset.
An MCP server that provides read access to your cloud storage providers, bank accounts and more.
Read-only finance MCP server. Connects your banks, cards, payment apps, books and payroll into one ledger for your AI. Can't move money. Open source (AGPL-3.0), free during beta.
Read-only MCP server for The Quiet Protocol's engines, benchmarks, proof, and business data.
Related MCP Servers
- AlicenseAqualityDmaintenanceRead-only MCP server for reconciling SimpleFIN bank data against Firefly III ledger, enabling audit and review workflows without mutation.1914 npm1MIT
- AlicenseNot gradedqualityCmaintenanceMCP server providing deterministic accounting tools for AI agents, including bank statement parsing, document classification, money math, and webhook verification.2Apache 2.0
- AlicenseAqualityDmaintenanceMCP server for deterministic numeric verification and spreadsheet auditing. Enables AI agents to verify math/finance claims, audit rows, and evaluate covenant rules with signed, independently checkable receipts.4MIT
- FlicenseBqualityDmaintenanceA secure, read-only MCP server that enables AI assistants to inspect transactions, vendor performance, wallet balances, and analytics through validated REST API calls.19-