job-search-mcp
job-search-mcp
一个面向 Claude Desktop 的个人求职助手。你让 Claude 找工作;它会搜索真实的招聘网站,为每个职位按 0–100 打分,评估它有多适合你,并显示在一个可排序的看板上,你可以一键筛选。一切都在本地运行——无需 API 密钥,无需注册账号。
这是一个 MCP 应用:一个 Claude Desktop 与之通信的小型服务器,外加一个直接渲染在聊天中的内联看板。
适用范围: 它针对美国软件工程岗位进行了调优(内置数据源和职位筛选器面向美国科技雇主)。其他领域或地区会返回稀疏的结果。
适用环境: 为 Claude Desktop 构建和测试。它也应该能在其他同时运行本地(stdio)服务器并渲染 MCP 应用界面的 MCP 客户端中工作,例如 VS Code(Copilot)、Cursor 或 Goose(未测试)。仅远程的客户端(ChatGPT、网页/移动端 Claude)无法启动本地服务器,因此无法使用。

它能做什么
搜索真实的招聘网站 —— LinkedIn 加上 8 个 ATS/职位数据源(Greenhouse、Lever、Ashby、Workday、SmartRecruiters、Hacker News、RemoteOK、Remotive),使用你的目标职位和地点。
为你给每个职位打分 —— Claude 阅读完整职位描述,给出 0–100 的匹配分数并附一行理由,综合考量你的技能、工作年限、职级匹配度和职位本身。
让你快速筛选 —— 在每张卡片上点击"申请/跳过",或批量操作("把 60 分以下的全部忽略")。
有记忆 —— 维护一个持续更新的候选清单、一个你已申请职位的追踪器,并且不会向你重复展示同一个职位(6 个月内)。
Related MCP server: Jobless MCP
设置
1. 告诉 Claude Desktop 这个应用。 打开你的配置文件:
Windows:
%APPDATA%\Claude\claude_desktop_config.jsonmacOS:
~/Library/Application Support/Claude/claude_desktop_config.json
在 "mcpServers" 下添加以下内容——npx 会为你获取已发布的包,无需克隆或构建:
{
"mcpServers": {
"job-search": {
"command": "npx",
"args": ["-y", "@servation/job-search-mcp"]
}
}
}2. 完全退出并重新打开 Claude Desktop(在 Windows 上,从系统托盘退出——仅关闭窗口是不够的)。就这样。
npm install
npm run build然后将配置指向此文件夹的 dist/main.js:
{
"mcpServers": {
"job-search": {
"command": "node",
"args": ["D:\\job-search-mcp\\dist\\main.js"]
}
}
}如何使用
直接和 Claude 对话即可。例如:
"保存我的档案" —— 先粘贴你的简历;Claude 会提取你的技能、职位、工作年限和地点,以便搜索和评分为你量身定制。
"找加州的高级后端工程师职位,本周发布的。"
"找申请人数少的职位" —— 添加 LinkedIn 的早期申请者筛选。
"把 60 分以下的全部忽略。"
"显示我已申请的职位。"
"重新给看板打分。"
典型的首次使用流程:保存你的档案 → "找工作" → Claude 打分并显示排序看板 → 你点击申请/跳过。
看板
看板是一个持续更新的候选清单。一个职位会一直留在上面,直到你申请(移入你的追踪器)或跳过(永久隐藏)。新的搜索会自动跳过已在看板上、追踪器中、已忽略的职位,以及过去 6 个月内展示过的职位——所以你永远不会连续两次看到同一个职位。
点击卡片上的已申请或跳过,或让 Claude 来做。在组件中完成筛选后,状态会持久保存。
工具(供参考)
你很少需要直接调用它们——Claude 会选择合适的工具——但以下是底层实现:
工具 | 功能 |
| 在招聘网站上搜索你的职位/地点(可选的筛选条件:职级、类型、发布时间、远程、薪资、申请人数、数据源)。 |
| Claude 为找到的职位按 0–100 打分评估匹配度(搜索后自动运行)。 |
| 显示排序看板(或你的 |
| 申请 / 跳过 / 保存单个职位(也是卡片按钮的功能)。 |
| 按分数或数据源批量筛选,例如把 60 分以下的全部忽略。 |
| 以文本形式列出当前看板(所有已打分职位 + 任何尚未打分的职位)。 |
| 你的追踪器——你已保存或已申请的职位。 |
| 从头重新为看板上的每个职位打分。 |
| 整理:清除未打分的遗留职位或整个看板(绝不会触碰已申请/已忽略的职位)。 |
| 保存你的简历档案(驱动搜索 + 打分)。 |
你的数据存放在哪里
你机器上的一个 JSON 文件——无云端,除了你搜索的招聘网站外,不会向任何地方发送任何内容:
已安装:
~/.job-search-mcp/jobs.json从源码运行:
./data/jobs.json可通过
JOB_SEARCH_MCP_DATA环境变量覆盖。
可选:LinkedIn Premium 模式
默认情况下,LinkedIn 使用其公开访客端点:无需登录,也没有账号登录风险,而且你仍然能获得完整的职位描述。请注意,即使以访客模式运行,自动化访问仍然来自你自己的 IP,并且违反 LinkedIn 的用户协议,因此请自行斟酌使用。如果你想要更丰富/Premium 的数据,可以通过在服务器配置中添加一个 env 块来使用你已登录的账号:
"job-search": {
"command": "npx",
"args": ["-y", "@servation/job-search-mcp"],
"env": {
"LINKEDIN_LI_AT": "<your li_at cookie>",
"LINKEDIN_JSESSIONID": "ajax:1234567890123456789"
}
}从已登录的 linkedin.com 标签页 → DevTools → Application → Cookies 中获取这两个 cookie。
⚠️ 注意: 这违反 LinkedIn 的服务条款,可能导致你的账号被标记或限制。Cookie 大约每月过期,而且 LinkedIn 的内部端点会随时更改,恕不另行通知。如果任何操作失败,它会自动回退到访客模式。不添加
env块以保持完全安全(仅访客模式)。
评分机制
Claude(宿主模型)直接进行评分——它阅读每份职位描述并判断匹配度,这具有良好的校准性。一个确定性公式(scoring.ts 中的 computeMatchScore)被保留作为回退方案,供那些运行较弱本地模型、倾向于给所有内容过高排名的用户使用;默认情况下不启用。
开发
npm run typecheck # type-check UI + server + harness
npm run build # build the UI bundle (vite single-file) + compile the server (tsc)
npm run serve:stdio # run the server from source (tsx) for local testing
npm run harness # work on the review UI without Claude Desktop (see harness/)
npm run canary # check every job source is still returning postings服务器仅支持 stdio(main.ts → server.ts);界面是一个 React 应用,打包为单个内联 HTML 文件(src/mcp-app.tsx → dist/mcp-app.html),由服务器作为 ui:// 资源提供。
开发界面时,使用 harness/。它是 Claude Desktop 的本地替代品,针对合成职位存储和真实的 ui/* 桥接渲染组件。在浏览器标签页中打开 dist/mcp-app.html 只会看到"正在连接……",而且最可能出问题的部分(链接打开、主题、挂载后的刷新)都由宿主中介,因此只有在另一端有宿主时才会显现。
如果某个职位数据源失效,npm run canary 会告诉你哪一个。公司 slug 经常失效,因此修复通常应放在 updateCompanyDirectoriesFromRegistry() 在运行时重新读取的远程 slug 注册表中,而不是发布新版本。
Available Tools
10 toolsbulk_statusBulk TriageA
Triage MANY board jobs at once by score and/or source — e.g. dismiss every scored job under 60. Applies status to all scanned (board) jobs matching the filters; the server selects them, so you do NOT need the individual job ids. below_score/above_score match SCORED jobs only (use clear_jobs for un-evaluated ones). Provide at least one filter. Returns text; call show_board afterward to display.
Inputs: status (required), notes, source, above_score, +1 more.
clear_jobsClear JobsA
Clear jobs from the review board. which='unscored' (default) removes only the not-yet-evaluated leftovers; which='all' clears the whole board. Saved/applied jobs (the tracker) and dismissed jobs are NOT affected — dismissed jobs still won't be re-sourced. Returns text; call show_board to display.
Inputs: which.
evaluate_jobsEvaluate JobsA
Score one or more sourced jobs for THIS candidate (holistic). YOU assign each job a 0-100 fit score directly, using the candidate summary (shown in find_jobs / whats_promising) and the job description. Weigh the must-have requirements, the candidate's actual skills and years of experience (a role demanding far more seniority/years than the candidate has should score LOWER), domain fit, and standout strengths. Calibrate and use the FULL range — do not bunch everything at 80+: 80-100 = strong fit and realistic; 60-79 = good with real gaps; 40-59 = partial/stretch (e.g. wrong level); below 40 = poor or wrong field. Pass { job_id, score, reason } per job. Use the EXACT bracketed ids; do not invent them. Score every job in one call. Returns text; call show_board afterward to display the ranked board.
Inputs: evaluations (required).
find_jobsFind JobsA
Source live jobs (LinkedIn + Greenhouse, Lever, Ashby, Workday, SmartRecruiters, Hacker News, RemoteOK, Remotive), dedup them, and return them as text with full descriptions. Uses the saved profile's roles/location by default; pass query/location/filters to override. This returns UNSCORED jobs and does NOT show a card UI — immediately score all of them with evaluate_jobs (a 0-100 fit score + reason each), then show_board once, to present the single ranked board. Do not just list them back.
Inputs: limit, query, remote, sort_by, +9 more.
rescore_boardRe-score BoardA
Re-score the WHOLE board from scratch — every job, INCLUDING ones that already have a score. Use this when the user asks to 'rescore'/'re-evaluate' the board or after the scoring approach changed. Returns all board jobs with their ids + descriptions; call evaluate_jobs with a FRESH { job_id, score, reason } for EACH (do not skip already-scored ones), then call show_board. Set include_saved=true to also re-score the saved/applied tracker.
Inputs: include_saved.
review_savedReview Saved JobsA
List the tracker as text: jobs marked saved (interested) or applied, with their status, score, and notes. Use this to see what you're tracking or have applied to. To DISPLAY the tracker widget, call show_board with view='saved'.
Inputs: no named parameters.
save_profileSave Resume ProfileA
Extract the candidate's profile from their resume text (which the user pastes in chat) and save it. Pull: full name; core technical skills; suggested target roles; preferred/search location; total years of professional experience (integer); and pass the resume text as rawText. The profile is used by find_jobs (target roles + location) and by scoring (years of experience drives the experience penalty), so set yearsOfExperience accurately. Merges with any existing profile.
Inputs: name, skills, rawText, targetRoles, +4 more.
set_statusSet Job StatusA
Triage a job by id: 'saved' (interested/tracking), 'applied' (also stamps the date), 'dismissed' (skip), or 'discovered' (undo back to the board). Moves it between the scanned/saved/dismissed lists and persists. Optionally attach notes. Called by the review UI; also usable directly.
Inputs: job_id (required), status (required), notes. Returns: jobs, view, count, scored, +1 more
show_boardShow BoardA
Display the job board (or the saved/applied tracker) as an interactive widget. Call this ONCE, at the end of a request, after doing the work (find_jobs/evaluate_jobs/bulk_status/clear_jobs/whats_promising are text-only and do NOT render). view='board' (default) shows the ranked board; view='saved' shows the tracker.
Inputs: view. Returns: jobs, view, count, scored, +1 more
whats_promisingWhat's PromisingA
List the current job board as text: every scored job (with its id) and any UNSCORED jobs with their ids + descriptions so you can evaluate them. To score the unscored ones, call evaluate_jobs with the EXACT bracketed ids shown. Use this to review jobs already found (find_jobs only returns brand-new ones). To re-score jobs that ALREADY have a score, use rescore_board instead. To DISPLAY the board widget to the user, call show_board (this returns text for you to act on).
Inputs: limit.
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 10 tools
Each tool has a clearly distinct purpose: bulk_status for bulk triage, clear_jobs for clearing, evaluate_jobs for scoring single jobs, find_jobs for sourcing, rescore_board for re-evaluating all, review_saved for tracker text, save_profile for profile, set_status for individual triage, show_board for display, and whats_promising for listing. No overlap.
All tool names follow a consistent verb_noun pattern in snake_case (e.g., bulk_status, evaluate_jobs, save_profile). The only minor deviation is 'whats_promising' using a contraction, but it still follows the pattern and is easily readable.
10 tools is well within the ideal 3-15 range for a focused job search server. Each tool addresses a necessary function without bloat, and the count feels natural for the domain.
The tool set covers the core workflow: sourcing (find_jobs), scoring (evaluate_jobs, rescore_board), triaging (bulk_status, set_status, clear_jobs), tracking (review_saved), display (show_board), and profile management (save_profile). Minor gaps like no explicit profile update tool exist, but save_profile merges, so it's almost complete.
Maintenance
Related MCP Connectors
AI job matching in your assistant: ranked offers, tailored resumes, application tracking.
Run your job search from your AI assistant. Career Rift gives every job its own resume, then works out the fit score, interview prep and cover letter from it, while your inbox keeps the job board current. Track applications, companies, notes, calendar events and profiles in plain language.
Tailor resumes, score ATS fit, find hiring leads, and send cold outreach, from your agent.
Job search for your AI: save jobs, research companies, keep stories, track applications.
Related MCP Servers
- AlicenseNot gradedqualityCmaintenanceEnables job search and scraping across multiple job boards (LinkedIn, Indeed, Glassdoor, etc.) with advanced filtering, directly from Claude Desktop or other MCP clients.6MIT
- AlicenseNot gradedqualityBmaintenancePersonalized job search inside Claude. Get ranked job matches based on your actual skills, not keywords.2MIT
- AlicenseAqualityDmaintenanceTransforms Claude into an AI job-hunting assistant that searches remote job boards, scores roles against your CV, generates tailored cover letters, and logs everything to a Notion tracker.11MIT
- AlicenseNot gradedqualityDmaintenanceEnables searching and evaluating job postings from LinkedIn and freehire.me directly through Claude Desktop. Provides tools to search jobs, fetch full posting details, and assess candidate fit using eligibility scans and a scoring rubric.MIT