"DeepSeek JSONL阅读能力" matching MCP tools:

• search_serversA

  MCP Aggregation Gateway — 22k+ MCP search & callA

  搜索 MCP Marketplace 平台目录中已收录的 MCP 服务（只读，共约 2.2 万个）。 【这是调用平台 MCP 的第一步】当你需要某个能力（如"查天气""查汇率""查热搜""查IP"）时， 先用本工具按关键词检索，拿到 server_id 与 endpoint_url，然后： - 用 list_server_tools(server_id=...) 查看该服务提供哪些工具； - 再用 call_server_tool(endpoint=..., tool_name=..., arguments={...}) 实际调用。 参数： query: 搜索关键词，中英文均可。支持多个关键词（空格分隔，AND 匹配）， 对名称(name)和描述(description)做模糊匹配。例如 "天气"、"weather"、"热搜 微博"。 limit: 返回条数上限，默认 10，最大 30。 返回：匹配的 MCP 列表，每条含： - id / name / slug / description（描述摘要） - endpoint_url: 远程 MCP 端点（若有；以 http/https 开头则 remote_direct=true，可直接远程调用） - remote_direct: 是否远程可直连 - alive / tools_count / checked_at: 平台能力验证状态（alive=1 表示最近验证存活，排在前面） 排序：验证存活(alive=1)且有远程端点的优先，其次按名称相关度。

  ConnectorNo auth

• zhini_list_tagsA

  知你AI助手｜多平台客户与客服数据 MCPA

  分页获取标签列表，用于在已知标签分组 ID 时列出该分组下的标签，并为客户搜索或历史会话搜索提供 tag_id。page 从 1 开始；page_size 最大 20。前置：可先调用 zhini_list_tag_groups 获取分组；如果用户只提供标签中文名，优先调用 zhini_search_tags 解析 tag_id。后续：拿到 tag_id 后可调用 zhini_search_customers 或 zhini_search_sessions。

  ConnectorNo auth

• zhini_search_customersA

  知你AI助手｜多平台客户与客服数据 MCPA

  搜索客户通讯录，用于根据客户名、手机号、微信号、客服、负责人、明确标签、渠道、性别、联系人类型或时间范围完整召回符合条件的客户 UID。用户只要求查询当天或某个时间范围内咨询过的客户时，默认查询当前授权账号关联的所有渠道范围，不代表当前调用客服本人接待或回复过；除非用户明确说“我处理过、我接待过、我回复过、归属于我”或指定某客服/渠道，否则不要自动传 kfid、pic_kfids 或 channel_id。用户要求查询某天或某段时间内的全部用户、包括会话已经结束/关闭的用户时，应使用本工具分页查询，不能用 zhini_list_active_sessions 代替。重要边界：自然语言中的“XXX 用户/客户”默认是业务语义或筛选条件，不应自动转换为 tag_id。可由名称、时间、渠道、客服等结构化字段表达的条件直接使用本工具；必须根据聊天内容判断的条件，应结合 zhini_search_sessions 和 zhini_fetch_messages 识别。只有用户明确要求某标签或上下文已有 tag_id 时才按标签筛选，且 tag_id 只覆盖已标注客户。用户给客户姓名/手机号/微信号时先用本工具召回候选客户；如果匹配多个客户，应让用户确认。用户明确给出标签名、渠道名、客服名时，应先分别调用 zhini_search_tags、zhini_list_channels、zhini_list_kefu 解析 ID。禁止空条件拉全量；page 从 0 开始；page_size 最大 20。后续：拿到 uid 后通常调用 zhini_get_customer_profile 或 zhini_fetch_messages。

  ConnectorNo auth

• analyze_shipA

  SateAIsA

  Sentinel-1 SAR画像（GRD）から船舶を検出します。 ■ 入力: 以下のどちらかで解析対象を指定。 (A) scene_id（Sentinel-1 GRDのシーンID）。ASF Search（https://search.asf.alaska.edu/#/）でFile TypeをGRDに絞って検索可能。 (B) polygon（WKT）+ date（基準日 YYYY-MM-DD）。該当範囲・基準日に近い1シーンを自動選定。date_direction で選択方向を指定可能。 ■ 処理: GPU推論による物体検出。推定所要時間は約30分。 ■ 出力: 検出された船舶のポリゴンを含むGeoJSON。 ■ フロー: ジョブ投入後 job_id が返される。check_job_status でポーリングし、completed 後に get_job_result で結果取得。 ■ 実行前に preview_analysis で解析範囲と消費クレジットを確認し、ユーザーの承認を得てください。ジョブは投入後にキャンセルできません。

  ConnectorOAuth

• calculate_payroll_batchA

  日本の給与・社会保険・労働法A

  Runs calculate_payslip for many employees in one call and returns the run totals: gross, employee deductions, net, and employer cost. Reach for this the moment more than two or three people are in play. A monthly payroll is not a sequence of unrelated questions — the employer share, the totals and the run id only mean anything across the whole run. Asking one employee at a time gives no total and no way to tell a retry from a second run. Put anything shared in defaults (prefecture, business_type, column) and let each row carry only what differs, which is usually pay and age. A row that cannot be computed comes back in errors with its index and id while the rest of the run completes — do not discard a whole payroll over one bad row. The reply carries a run_id derived from the route and the exact input, so sending the same payroll twice gives the same id. Nothing is stored, so a retry cannot double-count.

  ConnectorNo auth

• decide_regular_remuneration_batchA

  日本の給与・社会保険・労働法A

  Runs the annual 定時決定 for a whole payroll in one call, and reports which employees moved grade. 健康保険法第41条 puts every insured employee on the same schedule — the average of April, May and June pay, over the months with at least seventeen payment-basis days, applied from September to the following August. So June is the one month of the year when an office decides its entire payroll at once, and asking about one employee at a time is the wrong shape for the task. Reach for this the moment more than a couple of employees are in play. Each row returns the same judgement as decide_regular_remuneration, plus whether that person changed grade, which is what decides how much filing there is. Pass previous_remuneration to get that comparison; without it the answer is null rather than false, because "no grade to compare" and "did not move" are different facts. Pass acquired_on, left_on or revision_month and each row also says whether that employee is filed at all. 健康保険法第41条 leaves out anyone insured between 1 June and 1 July, anyone gone before the 1 July reference date, and anyone revised from July to September. The run totals to_file and not_required, which is the number of forms rather than the number of employees. A row that cannot be decided is returned in errors with its index and id, and the rest of the run still completes — do not discard a whole run over one bad row.

  ConnectorNo auth

• calculate_annual_costA

  日本の給与・社会保険・労働法A

  What one employee costs an employer over a year, bonuses included. Reach for this rather than multiplying a payslip by twelve, because the two do not agree once a bonus is paid. 健康保険法第45条 caps the standard bonus cumulatively across the year — 5,730,000 from 1 April to 31 March — so the same bonus costs a different amount depending on where it falls, and once the year is used up later bonuses carry no health premium. 厚生年金保険法第24条の4 caps at 1,500,000 per payment with no yearly total, so pension keeps charging where health has stopped. Pass bonuses in the order they are paid: the health allowance fills from the first one. Each row comes back with what was counted, whether it was cut, and how much of the year remains, so the answer can be explained rather than just quoted. Income tax here is the monthly figure times twelve. Bonus withholding is a separate calculation (calculate_bonus with include_tax) and the year-end adjustment is a separate tool (calculate_year_end_adjustment) — say so rather than presenting this as take-home pay for the year.

  ConnectorNo auth

• judge_worker_typeA

  日本の給与・社会保険・労働法A

  Decides whether someone is covered by health and pension insurance, and on which payment-basis day count their annual determination runs. Call this before decide_regular_remuneration or judge_monthly_revision whenever the person is anything other than plainly full-time. Those tools take a worker_type, and guessing it changes a real number: the determination counts months of seventeen payment-basis days for an ordinary employee and eleven for a 短時間労働者. Get the classification wrong and the answer is wrong with no sign of it. 健康保険法第3条第1項第9号 covers anyone whose weekly hours and monthly days reach three-quarters of a comparable full-time worker. Below that, four further tests decide it: twenty hours a week, 88,000 yen a month, not a student, and a workplace of at least fifty-one insured people. The engagement must also be expected to run past two months. The 88,000 figure leaves out overtime, bonuses, commuting and family allowances. Folding those in is the usual route to a wrong answer, so ask for 所定内賃金 specifically rather than total pay. What counts as a comparable full-time worker, and whether someone is a student for this purpose, are facts about the workplace and the person. Ask rather than assume; the tool applies the tests to what you pass and names any it could not evaluate.

  ConnectorNo auth

• get_statute_textA

  日本の給与・社会保険・労働法A

  Returns the full text of a Japanese statutory provision, as published by e-Gov. The judgement tools name the statute or notice their answer rests on, but not its words. Use this to quote the provision itself — a citation the reader can check beats a citation they have to take on trust, and Japanese payroll advice is routinely wrong in ways that only reading the article reveals. Only the provisions this API cites are bundled (about 28 across 8 laws); call it with no ref to list them. Abbreviations as practitioners write them (健保法43条, 厚年法81条の2, 徴収法11条), a missing 第, and paragraph-level references all resolve to the article. For anything outside this set, say so rather than reciting it from memory.

  ConnectorNo auth

• search_peopleA

  独行录 / opcmenuA

  【何时用】用户想找「人」时——「找能提供小程序代开发的主理人」「谁懂跨境电商供应链」「找人合作做 AI 出海产品」。搜的是主理人的供给侧（canOffer 能提供什么 + 昵称/介绍/身份标签），这是 OPC 之间撮合合作的刚需入口。 【机制】关键词 + 向量混合检索（RRF 融合），真人（已认领）梯队前置。结果含 canOffer / similarity / claimed。 【组合链】命中后 get_creator 看作品尽调 → start_conversation 开聊；对方若发过需求也可 contact_need 顺着需求接洽。搜「产品」用 search_products，搜「需求」用 search_needs。 【常见 pitfall】**不支持按手机号搜人**（隐私保护，服务端对手机号查询恒返回空）——用户给的是手机号时直接说明不支持，改问对方的昵称或能提供什么。

  ConnectorNo auth

• get_my_positioningA

  独行录 / opcmenuA

  【需要登录】【何时用】用户问「我现在到哪一步了」「接下来该干嘛」「帮我把这周能做的都做了」。返回五级主线的当前等级、六维画像、称号与总分，以及全部任务的完成态与 nextUp（最该做的三件事）。 【组合链·代办】nextUp 里每条任务都带 suggestedTool——**这是 agent 面相对 App 的关键差别：App 的 CTA 只能把人跳到那一屏让他自己动手，你能直接把这件事做完。** 对照表： · 写「我能提供什么」/ 一句话说清项目 → update_my_profile(canOffer)（partner.can_offer 与 funding.one_liner 两条任务判的就是 User.canOffer 的字数，分别 ≥30 / ≥20 字，写虚了过不了） · 融资资料、轮次金额、传 BP → set_my_role_profile(fundraising) · 邀请同行 → get_my_invite · 发需求（招兼职 / 找合伙人 / 找资源）→ create_need · 发布产品 → create_product · 归位产业链 → set_my_chain_position · 看看我的名片长什么样 → get_my_card；去回消息 → list_my_conversations 用户说「把这周能做的都做了」就真的一条条做完再汇报，别只念清单。 【口径/坑】 · **本工具会刷新你的定位快照**（写 PositioningState：算分快照 + auto 任务的完成戳），所以它不是纯读工具，别当免费接口循环调。只想看个大概用 get_my_brief。 · level.source='declared'（用户自报）**永远优先**于 inferred（LLM 读证据判的）/ observed（确定性兜底）。要改自报值走 set_my_role_profile(venture.stage)。 · basis / signals / confidence 是 LLM 给的判词，**转述它，别自己另判一个等级**，更别说「我觉得你其实已经到 XX 了」。 · 任务只增不减：达标那刻盖戳，之后数据回落也不打回未完成（用户不会莫名其妙掉级）。 · 默认**裁掉全部 39 条任务的培训正文（guide）**，否则一次调用几万 token。真要看某一条的正文：includeGuides=true 且必须同时给 taskKey，只取那一条。 · verify='manual' 的任务平台观测不到，要用 mark_positioning_task 自报打勾。

  ConnectorNo auth

• get_my_chat_openerA

  独行录 / opcmenuA

  【需要登录】【何时用】要改开场语之前先读现状；或者用户问「别人点找我聊聊时会收到什么」。 【组合链】读完 → 觉得该改就 set_my_chat_opener 写一句更像人说的。写之前先 get_my_card / get_my_products 读一遍他的「我能提供什么」和产品，写出来的话才有具体内容。 【口径/坑】opener=null 表示他没自定义，实际发出去的是 effective（全站默认那句）。这不是「没设置好」，默认那句本来就够用。

  ConnectorNo auth

• set_my_chat_openerA

  独行录 / opcmenuA

  【需要登录】【何时用】这是**纯文案活，正是 agent 最该替用户干的事**：读一遍他的「我能提供什么」和产品，替他写一句像真人说的开场白。 【这句话会自动发出去】别人点「找 TA 聊聊」时，服务端会**替你自动发出这一句**作为第一条消息（只在新建会话时发一次，不会刷屏）。所以它是一条**自动广播通道**，不是一条普通私信。 【怎么写】朴素、具体、不做当场能被戳穿的断言。三条硬规矩：① 不写「我懂你想要什么」这类你按按钮那刻根本不知道的话，对方回一句「那你说说」就穿帮；② 不用对仗押韵金句——顺口正是模板和 AI 文案的指纹；③ 说清「我从哪儿看到你的」，这是真的、可验证的，也天然给了对方话头。**不许出现任何「我是 AI 助手 / 自动发送」之类的标识**（产品口径：这就是他本人说的第一句话）。 【组合链】get_my_card（读 canOffer / 产品）→ 本工具写 → get_my_chat_opener 复核 → 之后 start_conversation 开的每个新会话都会自动带上它。传 null 或空串 = 恢复全站默认。 【口径/坑】 · 上限 120 字，超了报 chat_opener_too_long（400，终态，改短再提）。 · **不许夹联系方式和外链**：手机号 / 微信号 / QQ / 邮箱 / http 链接一律拒（chat_opener_has_contact）。这是自动广播面，放开就成了「加我微信卖课」的免费群发口。要换联系方式走双同意的 request_contact_exchange。 · 过敏感词闸（与私信同一把尺），命中报 content_rejected（400，终态，换写法，别原样重试）。

  ConnectorNo auth

• get_my_workA

  独行录 / opcmenuA

  【需要登录】查看独行录合作目标、待我做/我派出的任务、待回应合作邀请。任务保留目标与指派人，便于持续跟进。目标列表最多返回 limit 条并给总数，goalOffset 按 nextOffset 翻页（只影响目标列表）；任务 reachingLimit=true 时可能还有，按 goalId 调 list_collaboration_tasks 查看。邀请最多 50 条。读任务会幂等补齐周期任务的期次，故不是纯只读。不会读取或标记安排。下一步：get_collaboration_goal 看目标详情；set_collaboration_task_status 回报进展；respond_collaboration_invite 回应邀请；安排另用 get_my_dispatch。

  ConnectorNo auth

• zhini_list_tag_groupsA

  知你AI助手｜多平台客户与客服数据 MCPA

  获取标签分组，用于了解标签体系、解释标签归属，或在列出某分组标签前获取 tag_group_id。边界：只读字典工具，不新增、修改或删除标签分组。前置：无。后续：已知分组后可调用 zhini_list_tags；用户直接给标签名时通常优先调用 zhini_search_tags。

  ConnectorNo auth

• zhini_list_tag_groupsA

  知你客服A

  获取标签分组，用于了解标签体系、解释标签归属，或在列出某分组标签前获取 tag_group_id。边界：只读字典工具，不新增、修改或删除标签分组。前置：无。后续：已知分组后可调用 zhini_list_tags；用户直接给标签名时通常优先调用 zhini_search_tags。

  ConnectorNo auth

• data_queryA

  stagenth · 数据工具箱A

  查询 / 过滤 / 分组聚合数据文件，返回**实际数据行（JSON）**供 AI 直接分析（1 credit/次）。 支持 CSV/TSV/JSON/NDJSON/Parquet，两种用法： · 原始 SQL（表名固定 t）：sql="SELECT 商品, sum(销量) s FROM t GROUP BY 商品 ORDER BY s DESC LIMIT 5" · 结构化（不用写 SQL）：group_by=["地区"], measures=["销售额"], agg="sum", sort_by="销售额", descending=true, limit=10 SQL 仅允许单条只读 SELECT/WITH，禁止读文件/建表/联网。结果硬上限 1000 行，超出置 truncated=True。失败自动退款。 返回 {ok, format, mode, columns, total_rows, returned_rows, truncated, rows[]}。

  ConnectorNo auth

• ziwei_readingA

  FateStar ZiweiA

  调 FateStar「郑大钱」AI 命理师 —— 知识引擎 + 郑大钱人格深度解读命盘、回答命理问题 (感情/事业/财运/健康等)。解读会结合 FateStar 知识引擎与古籍检索上下文；排盘事实仍由确定性工具提供。【调用条件】仅在用户明确选择托管解读并确认可能扣积分后调用。普通命理问题默认先用免费排盘数据。⚠️ 付费: 按问题长度/模式扣积分 (短问免费), 需 FSFSKey 开发者 Key。【没 Key】引导用户去 https://www.fatestar.top 注册免费会员 → 做新手任务领积分 → 开发者中心拿 FSFSKey Key → 配进 MCP (环境变量 FATESTAR_API_KEY 或 apiKey 参数)。【积分用完 (402)】告知用户「已切回你自己的 LLM 解读, 请去 fatestar.top 充值或等北京时间 21:00 免费重置」, 然后改用 ziwei_chart 拿命盘数据 + 你自己解读。

  ConnectorNo auth

• dataset.exportA

  FlexOrch MCPA

  Download all records from a built dataset as text (Step 5 — final step). Returns the complete dataset content as a UTF-8 string directly in the response — no file download or separate URL needed. Call get_job_status after build_dataset and wait for status='completed' before calling this tool. Use the dataset_id from that completed response. Format guide: jsonl = LLM fine-tuning, rag = LangChain/LlamaIndex chunks, csv = spreadsheets, md = human-readable, xml = structured interchange. Binary formats (parquet, hf) cannot be returned via MCP — export them from the FlexOrch dashboard directly. Args: dataset_id: Dataset ID from the get_job_status completed build response. format: Text export format — jsonl, csv, json, md, xml, rag. Default: jsonl.

  ConnectorNo auth

• judge_annual_averageA

  日本の給与・社会保険・労働法A

  For work whose April-June happens to be its busiest or quietest quarter, where the ordinary calculation would fix a grade that is wrong for eleven months of the year. Available for 定時決定 (since April 2011) and 随時改定 (since October 2018). The 随時改定 figure is NOT a plain twelve-month average: it is the three-month average of FIXED pay plus the twelve-month average of NON-FIXED pay, so the two are supplied separately, and three distinct grade tests must all pass. Both routes need the employee's consent and require that the swing recurs every year for reasons inherent to the work — a one-off busy period does not qualify. Neither is something this tool can verify, so both are declared inputs and are echoed back in the response.

  ConnectorNo auth

• judge_annual_leaveA

  日本の給与・社会保険・労働法A

  Works out how many days of paid leave someone has been granted, and whether the employer still owes the five days it must direct. 労働基準法第39条 grants ten working days once six months of service are complete and attendance reaches eighty per cent of all working days, then adds one, two, four, six, eight and ten days in the years that follow. The ceiling everyone quotes as twenty is not in the article: it is the ten of the first grant plus the ten added from the sixth year. Someone under thirty hours a week working four days or fewer takes a smaller table from 施行規則第24条の3. Thirty hours is where it turns — at or above it the ordinary grant applies no matter how few days are worked, and treating such a person as part-time under-grants them. Ask for both the weekly days and the weekly hours; one without the other cannot settle it. Where ten or more days are granted, 第39条第7項 requires the employer to fix the timing of five of them within the year, and days the employee took of their own accord count toward it. A grant lapses two years after it is made (第115条), so one year carries over. The attendance figure is a question about the workplace: leave for a work injury, maternity, childcare and paid leave already taken all count as attendance. Ask for a rate that has been worked out rather than dividing days present by days in the year. Without one the tool reports the eighty per cent test as not judged rather than assuming it passed.

  ConnectorNo auth