AgentMesh Job Agent
jiyangnan/AgentMesh-JobAgent
Automação de hunting de empregos com IA para Boss直聘
Ver no GitHubAgent de busca de emprego AI-driven para plataforma Boss直聘 (Zhipin), automatizando o ciclo completo: extração de currículo, criação de perfil candidato, crawl de vagas, scoring inteligente e envio em lote de mensagens personalizadas. Projetado para ser controlado por agentes de IA como Claude Code e Cursor.
Instalação
npx convez-ai jiyangnan/AgentMesh-JobAgent1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Automação massiva de candidaturas em plataforma de empregos chinesa
- Scoring inteligente e matching de vagas com perfil do candidato
- Geração de mensagens personalizadas em lote com verificação de entrega
Veredito

Agent maduro e especializado para automação de job hunting, com arquitetura clara (cliente público + API privada). Exige licença para features de cloud; adequado para quem busca automação real em processo de recrutamento, não ferramenta genérica.
README
A tradução para o português deste README ainda está na fila.
Job Agent
🟣 Part of AgentMesh — see the ecosystem index (中文) for all related repos, the roadmap, and architecture.
![]()
![]()
AI-driven job hunting automation for Boss直聘 (Zhipin) — built to be controlled by your AI agent (Claude Code, OpenClaw, etc.) from chat.
Core loop: Resume → 36-field candidate profile (cloud) → Job crawl (local Chrome) → Match scoring (cloud) → Personalized greetings (cloud) → Batch send with delivery verification (local Chrome).
Architecture in one breath: Browser automation, your Boss cookie, and your resume original file always stay on your machine. Only stripped text and structured profile travel to our cloud API for matching/greeting generation.
This is the first product under the AgentMesh umbrella brand — a series of vertical AI agents for specific industries.
⚠️ Early access. Cloud features require a license key (request from the maintainer). M1 stage; the public release is intentionally low-key while we collect early feedback. See docs/progress.md for current status.
Architecture — two repos
Job Agent is split into two projects. You're looking at the client right now:
| Repo | Visibility | What it holds |
|---|---|---|
| Job Agent CLI (this repo) | Public · Apache 2.0 | Local browser automation, PDF/DOCX parsing, Boss API calls, IO, agent-friendly CLI surface. Calls the Cloud API for AI features. |
| Cloud API server | Private | Closed-source server holding our 36-field analysis prompt, match-scoring prompt, greeting-generation prompt, LLM provider keys, license/admin systems, the marketing landing site at jobagent.agentmesh360.com. This is our IP moat. |
If you're a new team member joining the project: ask the owner for access to the private server repo. It contains the operations runbook (docs/admin-runbook.md — how to issue licenses, deploy, debug), W1 external-resource checklist, and all server-side code. Don't try to find it by searching GitHub — the discoverable, public-facing surface is exactly this client repo plus jobagent.agentmesh360.com.
If you're an external user: you don't need access to the server repo. The Cloud API endpoint (api.jobagent.agentmesh360.com) and a license key are all you need; this client repo tells you everything else.
Features
- Resume Text Extraction — PDF / DOCX / TXT / Markdown → plain text
- Profile Management — Structured candidate profile (JSON), auto-merged into config
- Job Crawling — Fetch job listings from Boss Web API via real Chrome
- Smart Filtering — Exclude keywords, salary range, experience, degree
- AI Ranking — Rule-based scoring + optional LLM reranking
- Batch Greeting — Navigate job pages, click chat, fill message, send, verify delivery
- Audit Log — Full send history with success/failure tracking
- CDP Cross-Platform Driver — Real Chrome on macOS / Windows / Linux (not AppleScript-only)
- Passive Login Guide — Chrome auto-opens for login; agent notifies user; continues automatically
Requirements
- Python >= 3.11
- Google Chrome installed
- macOS / Windows / Linux
Installation
One-liner (recommended)
macOS / Linux (Terminal):
curl -fsSL https://raw.githubusercontent.com/jiyangnan/AgentMesh-JobAgent/main/scripts/install.sh | bash
Windows (PowerShell, regular user — no admin needed):
irm https://raw.githubusercontent.com/jiyangnan/AgentMesh-JobAgent/main/scripts/install.ps1 | iex
Both installers do the same six steps: check prereqs (Python 3.11+, git, Chrome), clone the repo to a hidden user-local directory, create an isolated venv, install the CLI, drop a jobagent shim on your PATH. After install, open a new terminal and you can run jobagent from anywhere.
Manual (developers / contributors)
git clone https://github.com/jiyangnan/AgentMesh-JobAgent.git
cd job-agent
python -m venv .venv
# macOS/Linux: source .venv/bin/activate
# Windows: .venv\Scripts\Activate.ps1
pip install -e ".[dev]"
🎁 Free vs License
CLI 本体永远免费 —— 抓岗位、管简历文件、自动化登录这些"agent 自己干不了"的事都给你。AI 那一段(简历理解、岗位匹配、招呼语)走云端,需要 license(M1 阶段免费申请,几小时内回)。
| Free · 本地工具版 | License · 完整云端 AI 版 | |
|---|---|---|
git clone 直接跑 | M1 阶段免费申请 | |
| 简历文本提取(PDF/DOCX → text) | ✅ | ✅ |
| Boss 岗位抓取(多页) | ✅ | ✅ |
| 独立 Chrome 自动化登录(cookie 留本地) | ✅ | ✅ |
| Profile JSON 手动管理(save/show/edit) | ✅ | ✅ |
| 投递审计日志 | ✅ | ✅ |
| AI 简历分析(36 字段招聘方画像) | ❌ | ✅ |
| AI 岗位匹配(5 维 + 5 档推荐 + 风险点) | ❌ | ✅ |
| AI 个性化招呼语(≤150 字 + 量化成果) | ❌ | ✅ |
| 批量发送 + 送达验证 | ❌ | ✅ |
| 云端算法持续迭代(无需更新客户端) | ❌ | ✅ |
| 未来产品免费试用(AgentMesh 矩阵) | ❌ | ✅ |
受 license 保护的命令:
jobagent resume analyze·jobs rank·greet preview·greet send·pipeline run。运行前会自动检查 license,没有就会停下来给你引导申请。Free 版可直接用:
resume extract、profile save/show/edit、jobs collect、login、doctor、init、greet audit。
🤖 For Agent Users (Claude Code / OpenClaw / Cursor / Codex …)
TL;DR: Don't paste this URL alone — paste docs/agent-onboarding.md instead.
This product is built to be driven by an AI agent (per strategy §1: users chat in IM, agents drive the CLI locally). But just handing your agent the GitHub link is risky — several steps need explicit handling:
| Concern | Why agent will trip |
|---|---|
| Cross-platform install one-liner | Windows users may be on Git Bash, can't run PowerShell irm |
jobagent login | Requires user to scan QR code in Chrome — agent must wait, not skip |
jobagent jobs collect throttling | CLI sleeps ~5s between pages to be courteous to the upstream API. Agent must NOT bypass with --page-delay 0 or parallel calls. |
jobagent greet send | Sends real messages on Boss直聘 — agent must ask for explicit confirmation before running |
| Quota / verification errors | Should stop and surface to user, not auto-retry with shorter delays |
🔑 The Boss-login step is the #1 thing agents fumble — here's what your agent MUST say
When the agent reaches jobagent login, it must read this exact prompt to you (the human) — verbatim, not paraphrased:
Agent → User:
接下来我要登录 Boss 直聘。我会启动一个独立的 Chrome 窗口(不是你日常用的那个 Chrome;这是为了隔离 cookie、保护你的隐私)。
请按以下步骤操作:
- 我马上会执行
jobagent login。一个新的 Chrome 窗口会自动弹出,地址栏是https://www.zhipin.com/。- 在那个新弹出的 Chrome 窗口里(不是你平时的 Chrome),用 Boss 直聘 App 扫码登录。
- 扫完后页面会跳转到你的 Boss 工作台,回到这里告诉我「登录好了」。
- 我收到你的"登录好了"后,会继续下一步。
⚠️ 不要关闭那个 Chrome 窗口;后续抓岗位和发招呼语都依赖它保持登录状态。
(5 分钟未扫码会超时,命令会自动退出——告诉我即可,我重新跑一遍。)
This explicit hand-off is the cornerstone of the IM→Agent→CLI flow. The agent should never silently run jobagent login and assume the user will figure out what to do. Always read the prompt above to the user, then run the command.
The canonical agent script
docs/agent-onboarding.md — copy the "快速指令" section, replace 4 placeholders, send to your agent. It addresses all the above pitfalls inline, including the literal login prompt above.
If you're a developer integrating with this product programmatically, the Claude Code skill at skills/claude-code/SKILL.md is the same workflow in Claude's preferred format — install once and the agent picks it up via natural-language trigger phrases.
Quick Start
TL;DR — which path should you pick?
Setup Recommended path You have a Cloud API license key (or can request one — currently free during M1) Path A: Cloud-driven (below). Best matches, best greetings, recruiter-perspective 36-field profile. You can't get a license today / want to evaluate first Path B: Pure-local (below). Works, but rule-based matching and template greetings — much weaker than cloud. You're a power user editing YAML configs by hand Path C: Manual config (further below). Requesting a license is fast and free during M1 — pick whichever channel fits:
- Application form: Apply here — 30-second structured form. Recommended.
- GitHub issue: open a license-request issue (public by default).
- Email:
hello@agentmesh360.com.All three reach the same queue; usually replied within a few hours.
Path A: Cloud-driven (recommended — uses Job Agent Cloud API)
# 1. One-time setup with your license key
jobagent init --key jba_live_xxxxxx
# Verifies connectivity, saves key to ~/.jobagent/credentials (mode 600)
# 2. Analyze resume (local extract → Cloud /v1/resume/analyze → save 36-field profile)
jobagent resume analyze --file resume.pdf --target-role "AI产品经理" --target-cities 深圳 杭州
# 3. Crawl jobs locally (uses your real Chrome + Boss cookie — never leaves your machine)
jobagent jobs collect --city 深圳 --query "AI产品经理" --output raw.json
# 4. Rank via cloud (sends profile + jobs to Cloud /v1/jobs/rank)
jobagent jobs rank --input raw.json --top 20 --output ranked.json
# 5. Preview personalized greetings (one Cloud /v1/greet/generate call per job)
jobagent greet preview --input ranked.json --limit 10 --output ready.json
# 6. Send (browser action stays local; the cloud_greeting from previous step is used)
jobagent greet send --input ready.json --limit 10
The Cloud API endpoint is https://api.jobagent.agentmesh360.com (override with
JOBAGENT_API_BASE env var). Browser automation, your Boss cookie, and your
resume original file always stay on your machine — only stripped text /
structured profile is sent to the cloud. See docs/product-strategy-20260509.md
for the full data flow.
Path B: Free-tier commands (no license)
Without a license you can still use the Free-tier commands as standalone tools — useful for evaluating the product locally before applying:
# Extract resume text (PDF/DOCX/TXT → text); useful even without a license
jobagent resume extract --file resume.pdf
# stderr prints a hint about the License version's upgrades
# Manage the profile JSON by hand (e.g. your own agent's LLM fills it)
jobagent profile save --data '{"basic":{...},"hardSkills":{...},...}'
jobagent profile show
jobagent profile edit
# Crawl jobs from Boss直聘 — landing in raw.json for inspection
jobagent jobs collect --city 深圳 --query "AI产品经理" --output raw.json
# Sanity-check the local environment
jobagent doctor env
# Read past send history (still works after license expiry)
jobagent greet audit
jobs rank, greet preview, greet send, resume analyze, pipeline run all require a license — they call the Cloud API and will exit with a friendly prompt if no license is configured. Apply for one here; M1 stage is free.
CLI Command Reference
# ── Login ──
jobagent login --check # Check Boss login status
jobagent login # Open Chrome login page, poll until logged in
# ── Resume / Profile (Step 1) ──
jobagent resume extract --file resume.pdf # Extract text from resume
jobagent profile save --data '{...}' # Save candidate profile JSON
jobagent profile show # Display current profile
# ── Job Discovery (Step 2) ──
jobagent jobs collect --city 深圳 --query "AI产品经理" [--output jobs.json]
jobagent jobs rank --input jobs.json --config config.yaml [--top 20] [--ai]
# ── Greeting (Step 3) ──
jobagent greet preview --input ranked.json [--limit 10]
jobagent greet send --input ranked.json [--limit 10]
jobagent greet audit [--recent 20]
# ── Diagnostics ──
jobagent doctor boss # Check Chrome / login / chat readiness
jobagent boss probe-send --job-url <url> --message "..."
jobagent boss verify-last-send
# ── Full Pipeline ──
jobagent pipeline run --config config.yaml
Agent Integration
This tool is designed to be orchestrated by an agent. Example workflow:
import subprocess, json
# Step 1: Extract resume
result = subprocess.run(
["jobagent", "resume", "extract", "--file", "resume.pdf"],
capture_output=True, text=True
)
resume_text = json.loads(result.stdout)["text"]
# Step 2: Agent uses its own LLM to build profile
profile = agent.llm.analyze_resume(resume_text)
# → {"years_experience": 10, "target_roles": ["AI产品经理"], ...}
# Step 3: Save profile
subprocess.run([
"jobagent", "profile", "save",
"--data", json.dumps(profile, ensure_ascii=False)
])
# Step 4: Run pipeline
subprocess.run(["jobagent", "pipeline", "run", "--config", "config.yaml"])
The product handles everything the agent cannot do: browser control, PDF parsing, Boss API calls, DOM automation. The agent handles everything it can do better: semantic understanding, LLM reasoning.
Configuration
config/config.yaml
platform: zhipin
crawler:
queries: ["AI产品经理", "高级产品经理"]
cities:
- name: 深圳
code: "101280600"
pages_per_query: 3
filter:
exclude_keywords: [销售, 客服, 保险]
max_salary_k: 0 # 0 = unlimited
require_degree_above: "大专"
greeter:
enabled: true
dry_run: false # true = preview only
template: |
您好,看了{company}的{job_name}岗位,我认为比较符合。
我有{years_experience}年{skills}相关经验,希望进一步沟通。
verify: true # verify [送达] mark after send
Profile auto-merge
If config.yaml has no candidate: section, Config.from_yaml() automatically loads ~/.jobagent/state/profile.json (created by jobagent profile save).
Priority: YAML explicit config > profile.json > defaults
Architecture
src/jobagent/
├── cli.py # Unified CLI entry
├── domain/ # Business logic
│ ├── models.py # Job, RankedJob, CandidateProfile, SendAttempt
│ ├── filter.py # FilterEngine (rule-based)
│ ├── ranking.py # RankingEngine (5-dim scoring + optional LLM rerank)
│ ├── greeter.py # GreeterEngine (preview / send_batch / verify)
│ ├── resume_parser.py # PDF/DOCX/TXT text extraction
│ └── profile_builder.py # Data validation (zero LLM)
├── drivers/boss/ # Browser automation
│ ├── cdp_driver.py # CDP implementation (cross-platform, primary)
│ ├── applescript_driver.py # AppleScript fallback (macOS only)
│ ├── chrome_manager.py # Chrome process management
│ ├── data_driver.py # Boss API fetching via Chrome XHR
│ └── __init__.py # create_driver() factory
├── application/ # Orchestration
│ ├── pipeline.py # 4-phase: crawl → filter → rank → greet
│ ├── doctor_boss.py # Session readiness checks
│ ├── probe_send.py # Single send test
│ └── verify_last_send.py
└── infra/ # Infrastructure
├── config.py # YAML loading + profile.json merge
├── state.py # State persistence paths
├── audit.py # Send audit log
└── exceptions.py # LoginRequiredError
Troubleshooting
| Problem | Solution |
|---|---|
| Chrome not found | Install Google Chrome. Paths checked: /Applications/Google Chrome.app (macOS), C:\Program Files\Google\Chrome\Application\chrome.exe (Windows), /usr/bin/google-chrome (Linux) |
| "登录态已失效" | Run jobagent login — Chrome will open automatically, log in, and the command continues |
| LLM not configured | This is expected. The product has no internal LLM. Your agent uses its own LLM. |
| Resume extraction fails (<100 chars) | The PDF may be a scanned image without OCR text layer. Try a text-based PDF. |
| Send fails with a verification challenge | The upstream redirected to a verify page. Pause for a while and retry. |
Development
# Run tests
pytest tests/
# Install in editable mode with dev deps
pip install -e ".[dev]"
Contact / Request a License
M1 license keys are free — we hand them out so we can track who's evaluating the product and collect feedback. Three ways to apply, pick whichever fits:
Option 1 — Application form (recommended)
Apply here (Tally form) — 30-second structured form, no GitHub account required. Goes to our queue; usually replied within a few hours via email.
Option 2 — GitHub Issue
Open a license-request issue — the template will collect what we need. Responses go back on the issue thread, so this option is public by default; choose Option 1 or 3 if you'd rather not have your request indexed.
Option 3 — Email
hello@agentmesh360.com — write a short note about what you're hunting and how you found us. We'll reply with a key.
What we look at
- Real-job-hunt use vs evaluation
- Which host agent you'd use (Claude Code / OpenClaw / Cursor / Codex / …)
- Where you heard about us
Nothing of the above is required to receive a key. M1 is free; we just want signal on who is reaching the product.
Privacy & Data
- 简历文件:始终保留在本机,不上传至任何云服务
- Cookie:Boss 直聘登录态仅在本机 Chrome 中使用,不离开本机
- AI 处理:简历提取文本和岗位信息通过加密通道发送至云端,由 DeepSeek 进行 AI 分析(匹配评分、招呼语生成)。数据按 DeepSeek 隐私政策 处理
- 服务端日志:仅记录 API 调用时间、接口名和 Token 用量,不记录简历内容
- 数据删除:如需删除使用记录,请发邮件至 hello@agentmesh360.com
Terms & Disclaimer
Job Agent 是独立第三方工具,与 Boss 直聘无合作关系。
- 服务边界:当前仅支持 macOS + Chrome + Boss 直聘,处于 Early Access 阶段,功能仍在迭代
- 平台风险:我们不承诺规避平台规则,账号被限制或封禁的风险由用户自行承担
- 用户责任:发送前请自行审核每条消息内容;禁止用于骚扰式海投
- 免责声明:不保证获得回复、面试或录用;对账号损失、数据丢失等不承担责任;产品按"现状"提供
完整服务条款:docs/marketing/reports/tos-disclaimer-minimum.md
License
MIT