2026ShippedTools
BOSS Client
boss-client
Turning BOSS Zhipin into a local, read-only data source that an agent can call.
- Role
- Design & development, solo
- Year
- 2026
- Status
- Shipped
- Stack
- TypeScriptMCPOpenAPISQLite
Job boards are one of the best public signals a salesperson has: what a company is hiring for, how many roles, how fast — it says a lot about what they are building and where the money goes. The problem is that BOSS Zhipin only exists as a web page. Search, pagination, job detail and company registration data all sit behind a login.
The problem
A script that hits the endpoints directly loses the session quickly. Two scripts sharing an account fight over the same browser profile, and one of them dies with “already running”. I did not want to hand the account to anyone or perform any write action — I just wanted to read public hiring data.
What I built
boss-client is a local read-only client. The core is one owner process: a systemd service is the only process allowed to hold the browser and profile. CLI, REST and MCP are just clients calling the local REST API; none of them opens a browser. A single-instance lock prevents two Chromium instances from writing to the same profile.
16 stable operations are defined once and generated into four contracts from the same registry: CLI commands, REST with an OpenAPI 3.1 document, MCP tools and a SKILL workflow. There is no “CLI has it, MCP doesn’t” gap.
Architecture notes
- Read-only is a design constraint: no messaging, no applications, no résumé edits. That is why it can stay running, and why there is no half-successful write to reason about.
- Low frequency: 12 requests per minute by default; SQLite caches jobs, companies, cities and search sessions, and only refreshes on TTL expiry.
- Login: the QR code is relayed to the terminal, so it runs headless. In practice one login lasts about ten days.
- Security: REST binds to 127.0.0.1 with a bearer token; profile, cache and logs are 600/700; captchas and cookies never reach long-term logs.
What I learned
The browser is the API. Once you accept that, the engineering worth doing is session lifetime, single ownership and caching — not scraping tricks.
One more thing: read-only gives the tool a boundary. It cannot apply for a job on my behalf, so it cannot make that decision wrongly either.